add files

data/coco-kp.yaml

@@ -0,0 +1,65 @@
+
+# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
+path: data/datasets/coco
+labels: kp_labels
+train: kp_labels/img_txt/train2017.txt
+val: kp_labels/img_txt/val2017.txt
+test: kp_labels/img_txt/test2017.txt
+
+train_annotations: annotations/person_keypoints_train2017.json
+val_annotations: annotations/person_keypoints_val2017.json
+test_annotations: annotations/image_info_test-dev2017.json
+
+pose_obj: True  # write pose object labels
+
+nc: 18  # number of classes (person class + 17 keypoint classes)
+num_coords: 34  # number of keypoint coordinates (x, y)
+
+# class names
+names: [ 'person', 'nose',
+         'left_eye', 'right_eye',
+         'left_ear', 'right_ear',
+         'left_shoulder', 'right_shoulder',
+         'left_elbow', 'right_elbow',
+         'left_wrist', 'right_wrist',
+         'left_hip', 'right_hip',
+         'left_knee', 'right_knee',
+         'left_ankle', 'right_ankle' ]
+
+kp_bbox: 0.05  # keypoint object size (normalized by longest img dim)
+kp_flip: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]  # for left-right keypoint flipping
+kp_left: [1, 3, 5, 7, 9, 11, 13, 15]  # left keypoints
+
+kp_names_short:
+  0: 'n'
+  1: 'ley'
+  2: 'rey'
+  3: 'lea'
+  4: 'rea'
+  5: 'ls'
+  6: 'rs'
+  7: 'lel'
+  8: 'rel'
+  9: 'lw'
+  10: 'rw'
+  11: 'lh'
+  12: 'rh'
+  13: 'lk'
+  14: 'rk'
+  15: 'la'
+  16: 'ra'
+
+# segments for plotting
+segments:
+  1: [5, 6]
+  2: [5, 11]
+  3: [11, 12]
+  4: [12, 6]
+  5: [5, 7]
+  6: [7, 9]
+  7: [6, 8]
+  8: [8, 10]
+  9: [11, 13]
+  10: [13, 15]
+  11: [12, 14]
+  12: [14, 16]

data/crowdpose.yaml

@@ -0,0 +1,62 @@
+
+# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
+path: data/datasets/crowdpose
+labels: kp_labels
+train: kp_labels/img_txt/trainval.txt
+val: kp_labels/img_txt/test.txt
+
+train_annotations: crowdpose_trainval.json
+val_annotations: crowdpose_test.json
+
+pose_obj: True  # write pose object labels
+
+nc: 15  # number of classes (person class + 14 keypoint classes)
+num_coords: 28 # number of keypoint coordinates (x, y)
+
+# class names
+names: [ 'person',
+         'left_shoulder', 'right_shoulder',
+         'left_elbow', 'right_elbow',
+         'left_wrist', 'right_wrist',
+         'left_hip', 'right_hip',
+         'left_knee', 'right_knee',
+         'left_ankle', 'right_ankle',
+         'head', 'neck']
+
+kp_bbox: 0.05  # keypoint object size (normalized by longest img dim)
+kp_flip: [1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 13]  # for left-right keypoint flipping
+kp_left: [0, 2, 4, 6, 8, 10]  # left keypoints
+
+kp_names_short:
+  0: 'ls'
+  1: 'rs'
+  2: 'lel'
+  3: 'rel'
+  4: 'lw'
+  5: 'rw'
+  6: 'lh'
+  7: 'rh'
+  8: 'lk'
+  9: 'rk'
+  10: 'la'
+  11: 'ra'
+  12: 'h'
+  13: 'n'
+
+# segments for plotting
+segments:
+  1: [0, 13]
+  2: [1, 13]
+  3: [0, 2]
+  4: [2, 4]
+  5: [1, 3]
+  6: [3, 5]
+  7: [0, 6]
+  8: [6, 7]
+  9: [7, 1]
+  10: [6, 8]
+  11: [8, 10]
+  12: [7, 9]
+  13: [9, 11]
+  14: [12, 13]
+

data/hyps/hyp.kp-p6.yaml

@@ -0,0 +1,36 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+# Hyperparameters for COCO training from scratch
+# python train.py --batch 32 --cfg yolov5m6.yaml --weights '' --data coco.yaml --img 1280 --epochs 300
+# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials
+
+lr0: 0.01  # initial learning rate (SGD=1E-2, Adam=1E-3)
+lrf: 0.2  # final OneCycleLR learning rate (lr0 * lrf)
+momentum: 0.937  # SGD momentum/Adam beta1
+weight_decay: 0.0005  # optimizer weight decay 5e-4
+warmup_epochs: 3.0  # warmup epochs (fractions ok)
+warmup_momentum: 0.8  # warmup initial momentum
+warmup_bias_lr: 0.1  # warmup initial bias lr
+box: 0.05  # box loss gain
+cls: 0.3  # cls loss gain
+cls_pw: 1.0  # cls BCELoss positive_weight
+obj: 0.7  # obj loss gain (scale with pixels)
+obj_pw: 1.0  # obj BCELoss positive_weight
+kp: 0.025  # kp loss gain
+iou_t: 0.20  # IoU training threshold
+anchor_t: 4.0  # anchor-multiple threshold
+# anchors: 3  # anchors per output layer (0 to ignore)
+fl_gamma: 0.0  # focal loss gamma (efficientDet default gamma=1.5)
+hsv_h: 0.015  # image HSV-Hue augmentation (fraction)
+hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.4  # image HSV-Value augmentation (fraction)
+degrees: 0.0  # image rotation (+/- deg)
+translate: 0.1  # image translation (+/- fraction)
+scale: 0.9  # image scale (+/- gain)
+shear: 0.0  # image shear (+/- deg)
+perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
+flipud: 0.0  # image flip up-down (probability)
+fliplr: 0.5  # image flip left-right (probability)
+mosaic: 1.0  # image mosaic (probability)
+mixup: 0.0  # image mixup (probability)
+copy_paste: 0.0  # segment copy-paste (probability)
+kp_bbox: 0.05

data/hyps/hyp.kp.yaml

@@ -0,0 +1,54 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+# Hyperparameters for COCO training from scratch
+# python train.py --batch 40 --cfg yolov5m.yaml --weights '' --data coco.yaml --img 640 --epochs 300
+# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials
+
+lr0: 0.01  # initial learning rate (SGD=1E-2, Adam=1E-3)
+lrf: 0.2  # final OneCycleLR learning rate (lr0 * lrf)
+momentum: 0.937  # SGD momentum/Adam beta1
+weight_decay: 0.0005  # optimizer weight decay 5e-4
+warmup_epochs: 3.0  # warmup epochs (fractions ok)
+warmup_momentum: 0.8  # warmup initial momentum
+warmup_bias_lr: 0.1  # warmup initial bias lr
+box: 0.05  # box loss gain
+cls: 0.5  # cls loss gain
+cls_pw: 1.0  # cls BCELoss positive_weight
+obj: 1.0  # obj loss gain (scale with pixels)
+obj_pw: 1.0  # obj BCELoss positive_weight
+kp: 0.05  # kp loss gain
+iou_t: 0.20  # IoU training threshold
+anchor_t: 4.0  # anchor-multiple threshold
+# anchors: 3  # anchors per output layer (0 to ignore)
+fl_gamma: 0.0  # focal loss gamma (efficientDet default gamma=1.5)
+hsv_h: 0.015  # image HSV-Hue augmentation (fraction)
+hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.4  # image HSV-Value augmentation (fraction)
+degrees: 0.0  # image rotation (+/- deg)
+translate: 0.2  # image translation (+/- fraction)
+scale: 0.8  # image scale (+/- gain)
+shear: 0.0  # image shear (+/- deg)
+perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
+flipud: 0.0  # image flip up-down (probability)
+fliplr: 0.5  # image flip left-right (probability)
+mosaic: 1.0  # image mosaic (probability)
+mixup: 0.0  # image mixup (probability)
+copy_paste: 0.0  # segment copy-paste (probability)
+kp_bbox: 0.05
+#kp_bbox:
+#- 0.026  # nose
+#- 0.025  # left eye
+#- 0.025  # right eye
+#- 0.035  # left ear
+#- 0.035  # right ear
+#- 0.079  # left shoulder
+#- 0.079  # right shoulder
+#- 0.072  # left elbow
+#- 0.072  # right elbow
+#- 0.062  # left wrist
+#- 0.062  # right wrist
+#- 0.107  # left hip
+#- 0.107  # right hip
+#- 0.087  # left knee
+#- 0.087  # right knee
+#- 0.089  # left ankle
+#- 0.089  # right ankle

data/scripts/download_models.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+# Example usage: bash data/scripts/download_models.sh
+
+gdown https://drive.google.com/uc?id=1hv0xwdbdf-Ym06Hpjy6wKyIklNmKmB99  # kapao_s_coco.pt
+gdown https://drive.google.com/uc?id=1B-QGa99n7ZrxkgKCQ1XHnm6j0Y-lvOQz  # kapao_m_coco.pt
+gdown https://drive.google.com/uc?id=1jYDfvRjhMoDf5xpMq1AuzdkFzUeLwt85  # kapao_l_coco.pt
+
+gdown https://drive.google.com/uc?id=1SmWwmqPwb_G6d9UPAFSUnWZ-4eaL9TTv  # kapao_s_crowdpose.pt
+gdown https://drive.google.com/uc?id=1IqrAk-gBdcfONrlIT6d4ndaDqnlFmT0r  # kapao_m_crowdpose.pt
+gdown https://drive.google.com/uc?id=146DW9ELzIBY2oDofPru446yLErxHrjJU  # kapao_l_crowdpose.pt

data/scripts/get_coco_kp.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+# Example usage: bash data/scripts/get_coco_kp.sh
+
+# Make dataset directories
+mkdir -p data/datasets/coco/images
+
+# Download/unzip annotations
+d='data/datasets/coco' # unzip directory
+f1='annotations_trainval2017.zip'
+f2='image_info_test2017.zip'
+url=http://images.cocodataset.org/annotations/
+for f in $f1 $f2; do
+  echo 'Downloading' $url$f '...'
+  curl -L $url$f -o $f && unzip -q $f -d $d && rm $f &
+done
+
+# Download/unzip images
+d='data/datasets/coco/images' # unzip directory
+url=http://images.cocodataset.org/zips/
+f1='train2017.zip' # 19G, 118k images
+f2='val2017.zip'   # 1G, 5k images
+f3='test2017.zip'  # 7G, 41k images (optional)
+for f in $f1 $f2 $f3; do
+  echo 'Downloading' $url$f '...'
+  curl -L $url$f -o $f && unzip -q $f -d $d && rm $f &
+done
+wait # finish background tasks

data/scripts/get_crowdpose.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+# Example usage: bash data/scripts/get_crowdpose.sh
+
+# Make dataset directories
+mkdir -p data/datasets/crowdpose
+
+gdown -O data/datasets/crowdpose/images.zip https://drive.google.com/uc?id=1VprytECcLtU4tKP32SYi_7oDRbw7yUTL
+gdown -O data/datasets/crowdpose/crowdpose_trainval.json https://drive.google.com/uc?id=13xScmTWqO6Y6m_CjiQ-23ptgX9sC-J9I
+gdown -O data/datasets/crowdpose/crowdpose_test.json https://drive.google.com/uc?id=1FUzRj-dPbL1OyBwcIX2BgFPEaY5Yrz7S
+unzip -q -d data/datasets/crowdpose data/datasets/crowdpose/images.zip
+rm data/datasets/crowdpose/images.zip

demos/flash_mob.py

@@ -0,0 +1,166 @@
+import sys
+from pathlib import Path
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[1].as_posix())  # add kapao/ to path
+
+import argparse
+from pytube import YouTube
+import os.path as osp
+from utils.torch_utils import select_device, time_sync
+from utils.general import check_img_size
+from utils.datasets import LoadImages
+from models.experimental import attempt_load
+import torch
+import cv2
+import numpy as np
+import yaml
+from tqdm import tqdm
+import imageio
+from val import run_nms, post_process_batch
+
+
+VIDEO_NAME = 'Crazy Uptown Funk Flashmob in Sydney for sydney domains campaign.mp4'
+URL = 'https://www.youtube.com/watch?v=2DiQUX11YaY&ab_channel=CrazyDomains'
+COLOR = (255, 0, 255)  # purple
+ALPHA = 0.5
+SEG_THICK = 3
+FPS_TEXT_SIZE = 2
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default='data/coco-kp.yaml')
+    parser.add_argument('--imgsz', type=int, default=1280)
+    parser.add_argument('--weights', default='kapao_s_coco.pt')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or cpu')
+    parser.add_argument('--half', action='store_true')
+    parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--no-kp-dets', action='store_true', help='do not use keypoint objects')
+    parser.add_argument('--conf-thres-kp', type=float, default=0.5)
+    parser.add_argument('--conf-thres-kp-person', type=float, default=0.2)
+    parser.add_argument('--iou-thres-kp', type=float, default=0.45)
+    parser.add_argument('--overwrite-tol', type=int, default=50)
+    parser.add_argument('--scales', type=float, nargs='+', default=[1])
+    parser.add_argument('--flips', type=int, nargs='+', default=[-1])
+    parser.add_argument('--display', action='store_true', help='display inference results')
+    parser.add_argument('--fps', action='store_true', help='display fps')
+    parser.add_argument('--gif', action='store_true', help='create fig')
+    parser.add_argument('--start', type=int, default=68, help='start time (s)')
+    parser.add_argument('--end', type=int, default=98, help='end time (s)')
+    args = parser.parse_args()
+
+    with open(args.data) as f:
+        data = yaml.safe_load(f)  # load data dict
+
+    # add inference settings to data dict
+    data['imgsz'] = args.imgsz
+    data['conf_thres'] = args.conf_thres
+    data['iou_thres'] = args.iou_thres
+    data['use_kp_dets'] = not args.no_kp_dets
+    data['conf_thres_kp'] = args.conf_thres_kp
+    data['iou_thres_kp'] = args.iou_thres_kp
+    data['conf_thres_kp_person'] = args.conf_thres_kp_person
+    data['overwrite_tol'] = args.overwrite_tol
+    data['scales'] = args.scales
+    data['flips'] = [None if f == -1 else f for f in args.flips]
+
+    if not osp.isfile(VIDEO_NAME):
+        yt = YouTube(URL)
+        # [print(s) for s in yt.streams]
+        stream = [s for s in yt.streams if s.itag == 136][0]  # 720p, non-progressive
+        print('Downloading squash demo video...')
+        stream.download()
+        print('Done.')
+
+    device = select_device(args.device, batch_size=1)
+    print('Using device: {}'.format(device))
+
+    model = attempt_load(args.weights, map_location=device)  # load FP32 model
+    half = args.half & (device.type != 'cpu')
+    if half:  # half precision only supported on CUDA
+        model.half()
+    stride = int(model.stride.max())  # model stride
+
+    imgsz = check_img_size(args.imgsz, s=stride)  # check image size
+    dataset = LoadImages('./{}'.format(VIDEO_NAME), img_size=imgsz, stride=stride, auto=True)
+
+    if device.type != 'cpu':
+        model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters())))  # run once
+
+    cap = dataset.cap
+    cap.set(cv2.CAP_PROP_POS_MSEC, args.start * 1000)
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    n = int(fps * (args.end - args.start))
+    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    gif_frames = []
+    video_name = 'flash_mob_inference_{}'.format(osp.splitext(args.weights)[0])
+
+    if not args.display:
+        writer = cv2.VideoWriter(video_name + '.mp4',
+                                 cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+        if not args.fps:  # tqdm might slows down inference
+            dataset = tqdm(dataset, desc='Writing inference video', total=n)
+
+    t0 = time_sync()
+    for i, (path, img, im0, _) in enumerate(dataset):
+        img = torch.from_numpy(img).to(device)
+        img = img.half() if half else img.float()  # uint8 to fp16/32
+        img = img / 255.0  # 0 - 255 to 0.0 - 1.0
+        if len(img.shape) == 3:
+            img = img[None]  # expand for batch dim
+
+        out = model(img, augment=True, kp_flip=data['kp_flip'], scales=data['scales'], flips=data['flips'])[0]
+        person_dets, kp_dets = run_nms(data, out)
+        bboxes, poses, _, _, _ = post_process_batch(data, img, [], [[im0.shape[:2]]], person_dets, kp_dets)
+
+        im0_copy = im0.copy()
+
+        # DRAW POSES
+        for j, (bbox, pose) in enumerate(zip(bboxes, poses)):
+            x1, y1, x2, y2 = bbox
+            size = ((x2 - x1) ** 2 + (y2 - y1) ** 2) ** 0.5
+            # if size < 450:
+            cv2.rectangle(im0_copy, (int(x1), int(y1)), (int(x2), int(y2)), COLOR, thickness=2)
+            for seg in data['segments'].values():
+                pt1 = (int(pose[seg[0], 0]), int(pose[seg[0], 1]))
+                pt2 = (int(pose[seg[1], 0]), int(pose[seg[1], 1]))
+                cv2.line(im0_copy, pt1, pt2, COLOR, SEG_THICK)
+        im0 = cv2.addWeighted(im0, ALPHA, im0_copy, 1 - ALPHA, gamma=0)
+
+        if i == 0:
+            t = time_sync() - t0
+        else:
+            t = time_sync() - t1
+
+        if args.fps:
+            s = FPS_TEXT_SIZE
+            cv2.putText(im0, '{:.1f} FPS'.format(1 / t), (5*s, 25*s),
+                        cv2.FONT_HERSHEY_SIMPLEX, s, (255, 255, 255), thickness=2*s)
+
+        if args.gif:
+            gif_frames.append(cv2.resize(im0, dsize=None, fx=0.375, fy=0.375)[:, :, [2, 1, 0]])
+        elif not args.display:
+            writer.write(im0)
+        else:
+            cv2.imshow('', im0)
+            cv2.waitKey(1)
+
+        t1 = time_sync()
+        if i == n - 1:
+            break
+
+    cv2.destroyAllWindows()
+    cap.release()
+    if not args.display:
+        writer.release()
+
+    if args.gif:
+        print('Saving GIF...')
+        with imageio.get_writer(video_name + '.gif', mode="I", fps=fps) as writer:
+            for idx, frame in tqdm(enumerate(gif_frames)):
+                writer.append_data(frame)
+
+
+

demos/general.py

@@ -0,0 +1,165 @@
+import sys
+from pathlib import Path
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[1].as_posix())  # add kapao/ to path
+
+import argparse
+from pytube import YouTube
+import os.path as osp
+from utils.torch_utils import select_device, time_sync
+from utils.general import check_img_size
+from utils.datasets import LoadImages
+from models.experimental import attempt_load
+import torch
+import cv2
+import numpy as np
+import yaml
+from tqdm import tqdm
+import imageio
+from val import run_nms, post_process_batch
+
+
+VIDEO_NAME = 'Crazy Uptown Funk Flashmob in Sydney for sydney domains campaign.mp4'
+URL = 'https://www.youtube.com/watch?v=2DiQUX11YaY&ab_channel=CrazyDomains'
+COLOR = (255, 0, 255)  # purple
+ALPHA = 0.5
+SEG_THICK = 3
+FPS_TEXT_SIZE = 2
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--vid', type=str, default='')
+    parser.add_argument('--data', type=str, default='data/coco-kp.yaml')
+    parser.add_argument('--imgsz', type=int, default=1280)
+    parser.add_argument('--weights', default='kapao_s_coco.pt')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or cpu')
+    parser.add_argument('--half', action='store_true')
+    parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--no-kp-dets', action='store_true', help='do not use keypoint objects')
+    parser.add_argument('--conf-thres-kp', type=float, default=0.5)
+    parser.add_argument('--conf-thres-kp-person', type=float, default=0.2)
+    parser.add_argument('--iou-thres-kp', type=float, default=0.45)
+    parser.add_argument('--overwrite-tol', type=int, default=50)
+    parser.add_argument('--scales', type=float, nargs='+', default=[1])
+    parser.add_argument('--flips', type=int, nargs='+', default=[-1])
+    parser.add_argument('--display', action='store_true', help='display inference results')
+    parser.add_argument('--fps', action='store_true', help='display fps')
+    parser.add_argument('--gif', action='store_true', help='create fig')
+    parser.add_argument('--start', type=int, default=68, help='start time (s)')
+    parser.add_argument('--end', type=int, default=98, help='end time (s)')
+    args = parser.parse_args()
+
+    with open(args.data) as f:
+        data = yaml.safe_load(f)  # load data dict
+
+    # add inference settings to data dict
+    data['imgsz'] = args.imgsz
+    data['conf_thres'] = args.conf_thres
+    data['iou_thres'] = args.iou_thres
+    data['use_kp_dets'] = not args.no_kp_dets
+    data['conf_thres_kp'] = args.conf_thres_kp
+    data['iou_thres_kp'] = args.iou_thres_kp
+    data['conf_thres_kp_person'] = args.conf_thres_kp_person
+    data['overwrite_tol'] = args.overwrite_tol
+    data['scales'] = args.scales
+    data['flips'] = [None if f == -1 else f for f in args.flips]
+
+    if not osp.isfile(VIDEO_NAME):
+        yt = YouTube(URL)
+        # [print(s) for s in yt.streams]
+        stream = [s for s in yt.streams if s.itag == 136][0]  # 720p, non-progressive
+        print('Downloading squash demo video...')
+        stream.download()
+        print('Done.')
+
+    device = select_device(args.device, batch_size=1)
+    print('Using device: {}'.format(device))
+
+    model = attempt_load(args.weights, map_location=device)  # load FP32 model
+    half = args.half & (device.type != 'cpu')
+    if half:  # half precision only supported on CUDA
+        model.half()
+    stride = int(model.stride.max())  # model stride
+
+    imgsz = check_img_size(args.imgsz, s=stride)  # check image size
+    dataset = LoadImages('./{}'.format(args.vid), img_size=imgsz, stride=stride, auto=True)
+
+    if device.type != 'cpu':
+        model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters())))  # run once
+
+    cap = dataset.cap
+    cap.set(cv2.CAP_PROP_POS_MSEC, args.start * 1000)
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    n = int(fps * (args.end - args.start))
+    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    gif_frames = []
+    video_name = 'flash_mob_inference_{}'.format(osp.splitext(args.weights)[0])
+
+    if not args.display:
+        writer = cv2.VideoWriter(video_name + '.mp4',
+                                 cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+        if not args.fps:  # tqdm might slows down inference
+            dataset = tqdm(dataset, desc='Writing inference video', total=n)
+
+    t0 = time_sync()
+    for i, (path, img, im0, _) in enumerate(dataset):
+        img = torch.from_numpy(img).to(device)
+        img = img.half() if half else img.float()  # uint8 to fp16/32
+        img = img / 255.0  # 0 - 255 to 0.0 - 1.0
+        if len(img.shape) == 3:
+            img = img[None]  # expand for batch dim
+
+        out = model(img, augment=True, kp_flip=data['kp_flip'], scales=data['scales'], flips=data['flips'])[0]
+        person_dets, kp_dets = run_nms(data, out)
+        bboxes, poses, _, _, _ = post_process_batch(data, img, [], [[im0.shape[:2]]], person_dets, kp_dets)
+
+        im0_copy = im0.copy()
+
+        # DRAW POSES
+        for j, (bbox, pose) in enumerate(zip(bboxes, poses)):
+            x1, y1, x2, y2 = bbox
+            size = ((x2 - x1) ** 2 + (y2 - y1) ** 2) ** 0.5
+            # if size < 450:
+            cv2.rectangle(im0_copy, (int(x1), int(y1)), (int(x2), int(y2)), COLOR, thickness=2)
+            for seg in data['segments'].values():
+                pt1 = (int(pose[seg[0], 0]), int(pose[seg[0], 1]))
+                pt2 = (int(pose[seg[1], 0]), int(pose[seg[1], 1]))
+                cv2.line(im0_copy, pt1, pt2, COLOR, SEG_THICK)
+        im0 = cv2.addWeighted(im0, ALPHA, im0_copy, 1 - ALPHA, gamma=0)
+
+        if i == 0:
+            t = time_sync() - t0
+        else:
+            t = time_sync() - t1
+
+        if args.fps:
+            s = FPS_TEXT_SIZE
+            cv2.putText(im0, '{:.1f} FPS'.format(1 / t), (5*s, 25*s),
+                        cv2.FONT_HERSHEY_SIMPLEX, s, (255, 255, 255), thickness=2*s)
+
+        if args.gif:
+            gif_frames.append(cv2.resize(im0, dsize=None, fx=0.375, fy=0.375)[:, :, [2, 1, 0]])
+        elif not args.display:
+            writer.write(im0)
+        else:
+            cv2.imshow('', im0)
+            cv2.waitKey(1)
+
+        t1 = time_sync()
+        if i == n - 1:
+            break
+
+    cv2.destroyAllWindows()
+    cap.release()
+    if not args.display:
+        writer.release()
+
+    if args.gif:
+        print('Saving GIF...')
+        with imageio.get_writer(video_name + '.gif', mode="I", fps=fps) as writer:
+            for idx, frame in tqdm(enumerate(gif_frames)):
+                writer.append_data(frame)
+

demos/squash.py

@@ -0,0 +1,243 @@
+import sys
+from pathlib import Path
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[1].as_posix())  # add kapao/ to path
+
+import argparse
+from pytube import YouTube
+import os.path as osp
+from utils.torch_utils import select_device, time_sync
+from utils.general import check_img_size
+from utils.datasets import LoadImages
+from models.experimental import attempt_load
+import torch
+import cv2
+import numpy as np
+import yaml
+from tqdm import tqdm
+import imageio
+from val import run_nms, post_process_batch
+
+
+VIDEO_NAME = 'Squash MegaRally 176 ReDux - Slow Mo Edition.mp4'
+URL = 'https://www.youtube.com/watch?v=Dy62-eTNvY4&ab_channel=PSASQUASHTV'
+
+GRAY = (200, 200, 200)
+CROWD_THRES = 450  # max bbox size for crowd classification
+CROWD_ALPHA = 0.5
+CROWD_KP_SIZE = 2
+CROWD_KP_THICK = 2
+CROWD_SEG_THICK = 2
+
+BLUE = (245, 140, 66)
+ORANGE = (66, 140, 245)
+PLAYER_ALPHA_BOX = 0.85
+PLAYER_ALPHA_POSE = 0.3
+PLAYER_KP_SIZE = 4
+PLAYER_KP_THICK = 4
+PLAYER_SEG_THICK = 4
+FPS_TEXT_SIZE = 3
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default='data/coco-kp.yaml')
+    parser.add_argument('--imgsz', type=int, default=1280)
+    parser.add_argument('--weights', default='kapao_s_coco.pt')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or cpu')
+    parser.add_argument('--half', action='store_true')
+    parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--no-kp-dets', action='store_true', help='do not use keypoint objects')
+    parser.add_argument('--conf-thres-kp', type=float, default=0.5)
+    parser.add_argument('--conf-thres-kp-person', type=float, default=0.2)
+    parser.add_argument('--iou-thres-kp', type=float, default=0.45)
+    parser.add_argument('--overwrite-tol', type=int, default=50)
+    parser.add_argument('--scales', type=float, nargs='+', default=[1])
+    parser.add_argument('--flips', type=int, nargs='+', default=[-1])
+    parser.add_argument('--display', action='store_true', help='display inference results')
+    parser.add_argument('--fps', action='store_true', help='display fps')
+    parser.add_argument('--gif', action='store_true', help='create fig')
+    parser.add_argument('--start', type=int, default=20, help='start time (s)')
+    parser.add_argument('--end', type=int, default=80, help='end time (s)')
+    args = parser.parse_args()
+
+    with open(args.data) as f:
+        data = yaml.safe_load(f)  # load data dict
+
+    # add inference settings to data dict
+    data['imgsz'] = args.imgsz
+    data['conf_thres'] = args.conf_thres
+    data['iou_thres'] = args.iou_thres
+    data['use_kp_dets'] = not args.no_kp_dets
+    data['conf_thres_kp'] = args.conf_thres_kp
+    data['iou_thres_kp'] = args.iou_thres_kp
+    data['conf_thres_kp_person'] = args.conf_thres_kp_person
+    data['overwrite_tol'] = args.overwrite_tol
+    data['scales'] = args.scales
+    data['flips'] = [None if f == -1 else f for f in args.flips]
+
+    if not osp.isfile(VIDEO_NAME):
+        yt = YouTube(URL)
+        # [print(s) for s in yt.streams]
+        stream = [s for s in yt.streams if s.itag == 137][0]  # 1080p, non-progressive
+        print('Downloading squash demo video...')
+        stream.download()
+        print('Done.')
+
+    device = select_device(args.device, batch_size=1)
+    print('Using device: {}'.format(device))
+
+    model = attempt_load(args.weights, map_location=device)  # load FP32 model
+    half = args.half & (device.type != 'cpu')
+    if half:  # half precision only supported on CUDA
+        model.half()
+    stride = int(model.stride.max())  # model stride
+
+    imgsz = check_img_size(args.imgsz, s=stride)  # check image size
+    dataset = LoadImages('./{}'.format(VIDEO_NAME), img_size=imgsz, stride=stride, auto=True)
+
+    if device.type != 'cpu':
+        model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters())))  # run once
+
+    cap = dataset.cap
+    cap.set(cv2.CAP_PROP_POS_MSEC, args.start * 1000)
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    n = int(fps * (args.end - args.start))
+    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    gif_frames = []
+    video_name = 'squash_inference_{}'.format(osp.splitext(args.weights)[0])
+
+    if not args.display:
+        writer = cv2.VideoWriter(video_name + '.mp4',
+                                 cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+        if not args.fps:  # tqdm might slows down inference
+            dataset = tqdm(dataset, desc='Writing inference video', total=n)
+
+    t0 = time_sync()
+    for i, (path, img, im0, _) in enumerate(dataset):
+        img = torch.from_numpy(img).to(device)
+        img = img.half() if half else img.float()  # uint8 to fp16/32
+        img = img / 255.0  # 0 - 255 to 0.0 - 1.0
+        if len(img.shape) == 3:
+            img = img[None]  # expand for batch dim
+
+        out = model(img, augment=True, kp_flip=data['kp_flip'], scales=data['scales'], flips=data['flips'])[0]
+        person_dets, kp_dets = run_nms(data, out)
+        bboxes, poses, _, _, _ = post_process_batch(data, img, [], [[im0.shape[:2]]], person_dets, kp_dets)
+
+        bboxes = np.array(bboxes)
+        poses = np.array(poses)
+
+        im0_copy = im0.copy()
+        player_idx = []
+
+        # DRAW CROWD POSES
+        for j, (bbox, pose) in enumerate(zip(bboxes, poses)):
+            x1, y1, x2, y2 = bbox
+            size = ((x2 - x1) ** 2 + (y2 - y1) ** 2) ** 0.5
+            if size < CROWD_THRES:
+                cv2.rectangle(im0_copy, (int(x1), int(y1)), (int(x2), int(y2)), GRAY, thickness=2)
+                for x, y, _ in pose[:5]:
+                    cv2.circle(im0_copy, (int(x), int(y)), CROWD_KP_SIZE, GRAY, CROWD_KP_THICK)
+                for seg in data['segments'].values():
+                    pt1 = (int(pose[seg[0], 0]), int(pose[seg[0], 1]))
+                    pt2 = (int(pose[seg[1], 0]), int(pose[seg[1], 1]))
+                    cv2.line(im0_copy, pt1, pt2, GRAY, CROWD_SEG_THICK)
+            else:
+                player_idx.append(j)
+        im0 = cv2.addWeighted(im0, CROWD_ALPHA, im0_copy, 1 - CROWD_ALPHA, gamma=0)
+
+        # DRAW PLAYER POSES
+        player_bboxes = bboxes[player_idx][:2]
+        player_poses = poses[player_idx][:2]
+
+        def draw_player_poses(im0, missing=-1):
+            for j, (bbox, pose, color) in enumerate(zip(
+                    player_bboxes[[orange_player, blue_player]],
+                    player_poses[[orange_player, blue_player]],
+                    [ORANGE, BLUE])):
+                if j == missing:
+                    continue
+                im0_copy = im0.copy()
+                x1, y1, x2, y2 = bbox
+                cv2.rectangle(im0_copy, (int(x1), int(y1)), (int(x2), int(y2)), color, thickness=-1)
+                im0 = cv2.addWeighted(im0, PLAYER_ALPHA_BOX, im0_copy, 1 - PLAYER_ALPHA_BOX, gamma=0)
+                im0_copy = im0.copy()
+                for x, y, _ in pose:
+                    cv2.circle(im0_copy, (int(x), int(y)), PLAYER_KP_SIZE, color, PLAYER_KP_THICK)
+                for seg in data['segments'].values():
+                    pt1 = (int(pose[seg[0], 0]), int(pose[seg[0], 1]))
+                    pt2 = (int(pose[seg[1], 0]), int(pose[seg[1], 1]))
+                    cv2.line(im0_copy, pt1, pt2, color, PLAYER_SEG_THICK)
+                im0 = cv2.addWeighted(im0, PLAYER_ALPHA_POSE, im0_copy, 1 - PLAYER_ALPHA_POSE, gamma=0)
+            return im0
+
+        if i == 0:
+            # orange player on left at start
+            orange_player = np.argmin(player_bboxes[:, 0])
+            blue_player = int(not orange_player)
+            im0 = draw_player_poses(im0)
+        else:
+            # simple player tracking based on frame-to-frame pose difference
+            dist = []
+            for pose in poses_last:
+                dist.append(np.mean(np.linalg.norm(player_poses[0, :, :2] - pose[:, :2], axis=-1)))
+            if np.argmin(dist) == 0:
+                orange_player = 0
+            else:
+                orange_player = 1
+            blue_player = int(not orange_player)
+
+            # if only one player detected, find which player is missing
+            missing = -1
+            if len(player_poses) == 1:
+                if orange_player == 0:  # missing blue player
+                    player_poses = np.concatenate((player_poses, poses_last[1:]), axis=0)
+                    player_bboxes = np.concatenate((player_bboxes, bboxes_last[1:]), axis=0)
+                    missing = 1
+                else:  # missing orange player
+                    player_poses = np.concatenate((player_poses, poses_last[:1]), axis=0)
+                    player_bboxes = np.concatenate((player_bboxes, bboxes_last[:1]), axis=0)
+                    missing = 0
+            im0 = draw_player_poses(im0, missing)
+
+        bboxes_last = player_bboxes[[orange_player, blue_player]]
+        poses_last = player_poses[[orange_player, blue_player]]
+
+        if i == 0:
+            t = time_sync() - t0
+        else:
+            t = time_sync() - t1
+
+        if args.fps:
+            s = FPS_TEXT_SIZE
+            cv2.putText(im0, '{:.1f} FPS'.format(1 / t), (5*s, 25*s),
+                        cv2.FONT_HERSHEY_SIMPLEX, s, (255, 255, 255), thickness=2*s)
+
+        if args.gif:
+            gif_frames.append(cv2.resize(im0, dsize=None, fx=0.25, fy=0.25)[:, :, [2, 1, 0]])
+        elif not args.display:
+            writer.write(im0)
+        else:
+            cv2.imshow('', cv2.resize(im0, dsize=None, fx=0.5, fy=0.5))
+            cv2.waitKey(1)
+
+        t1 = time_sync()
+        if i == n - 1:
+            break
+
+    cv2.destroyAllWindows()
+    cap.release()
+    if not args.display:
+        writer.release()
+
+    if args.gif:
+        print('Saving GIF...')
+        with imageio.get_writer(video_name + '.gif', mode="I", fps=fps) as writer:
+            for idx, frame in tqdm(enumerate(gif_frames)):
+                writer.append_data(frame)
+
+
+

models/common.py

@@ -0,0 +1,452 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Common modules
+"""
+
+import logging
+import math
+import warnings
+from copy import copy
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import requests
+import torch
+import torch.nn as nn
+from PIL import Image
+from torch.cuda import amp
+
+from utils.datasets import exif_transpose, letterbox
+from utils.general import colorstr, increment_path, is_ascii, make_divisible, non_max_suppression, save_one_box, \
+    scale_coords, xyxy2xywh
+from utils.plots import Annotator, colors
+from utils.torch_utils import time_sync
+
+LOGGER = logging.getLogger(__name__)
+
+
+def autopad(k, p=None):  # kernel, padding
+    # Pad to 'same'
+    if p is None:
+        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
+    return p
+
+
+class Conv(nn.Module):
+    # Standard convolution
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
+        self.bn = nn.BatchNorm2d(c2)
+        self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
+
+    def forward(self, x):
+        return self.act(self.bn(self.conv(x)))
+
+    def forward_fuse(self, x):
+        return self.act(self.conv(x))
+
+
+class DWConv(Conv):
+    # Depth-wise convolution class
+    def __init__(self, c1, c2, k=1, s=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
+
+
+class TransformerLayer(nn.Module):
+    # Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance)
+    def __init__(self, c, num_heads):
+        super().__init__()
+        self.q = nn.Linear(c, c, bias=False)
+        self.k = nn.Linear(c, c, bias=False)
+        self.v = nn.Linear(c, c, bias=False)
+        self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
+        self.fc1 = nn.Linear(c, c, bias=False)
+        self.fc2 = nn.Linear(c, c, bias=False)
+
+    def forward(self, x):
+        x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x
+        x = self.fc2(self.fc1(x)) + x
+        return x
+
+
+class TransformerBlock(nn.Module):
+    # Vision Transformer https://arxiv.org/abs/2010.11929
+    def __init__(self, c1, c2, num_heads, num_layers):
+        super().__init__()
+        self.conv = None
+        if c1 != c2:
+            self.conv = Conv(c1, c2)
+        self.linear = nn.Linear(c2, c2)  # learnable position embedding
+        self.tr = nn.Sequential(*[TransformerLayer(c2, num_heads) for _ in range(num_layers)])
+        self.c2 = c2
+
+    def forward(self, x):
+        if self.conv is not None:
+            x = self.conv(x)
+        b, _, w, h = x.shape
+        p = x.flatten(2).unsqueeze(0).transpose(0, 3).squeeze(3)
+        return self.tr(p + self.linear(p)).unsqueeze(3).transpose(0, 3).reshape(b, self.c2, w, h)
+
+
+class Bottleneck(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c2, 3, 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class BottleneckCSP(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
+        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
+        self.act = nn.LeakyReLU(0.1, inplace=True)
+        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.cv3(self.m(self.cv1(x)))
+        y2 = self.cv2(x)
+        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
+
+
+class C3(nn.Module):
+    # CSP Bottleneck with 3 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1)  # act=FReLU(c2)
+        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+        # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
+
+    def forward(self, x):
+        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))
+
+
+class C3TR(C3):
+    # C3 module with TransformerBlock()
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)
+        self.m = TransformerBlock(c_, c_, 4, n)
+
+
+class C3SPP(C3):
+    # C3 module with SPP()
+    def __init__(self, c1, c2, k=(5, 9, 13), n=1, shortcut=True, g=1, e=0.5):
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)
+        self.m = SPP(c_, c_, k)
+
+
+class C3Ghost(C3):
+    # C3 module with GhostBottleneck()
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[GhostBottleneck(c_, c_) for _ in range(n)])
+
+
+class SPP(nn.Module):
+    # Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729
+    def __init__(self, c1, c2, k=(5, 9, 13)):
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
+        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
+
+    def forward(self, x):
+        x = self.cv1(x)
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning
+            return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
+
+
+class SPPF(nn.Module):
+    # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
+    def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * 4, c2, 1, 1)
+        self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
+
+    def forward(self, x):
+        x = self.cv1(x)
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning
+            y1 = self.m(x)
+            y2 = self.m(y1)
+            return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
+
+
+class Focus(nn.Module):
+    # Focus wh information into c-space
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
+        # self.contract = Contract(gain=2)
+
+    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
+        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
+        # return self.conv(self.contract(x))
+
+
+class GhostConv(nn.Module):
+    # Ghost Convolution https://github.com/huawei-noah/ghostnet
+    def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
+        super().__init__()
+        c_ = c2 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, k, s, None, g, act)
+        self.cv2 = Conv(c_, c_, 5, 1, None, c_, act)
+
+    def forward(self, x):
+        y = self.cv1(x)
+        return torch.cat([y, self.cv2(y)], 1)
+
+
+class GhostBottleneck(nn.Module):
+    # Ghost Bottleneck https://github.com/huawei-noah/ghostnet
+    def __init__(self, c1, c2, k=3, s=1):  # ch_in, ch_out, kernel, stride
+        super().__init__()
+        c_ = c2 // 2
+        self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1),  # pw
+                                  DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(),  # dw
+                                  GhostConv(c_, c2, 1, 1, act=False))  # pw-linear
+        self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
+                                      Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
+
+    def forward(self, x):
+        return self.conv(x) + self.shortcut(x)
+
+
+class Contract(nn.Module):
+    # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        b, c, h, w = x.size()  # assert (h / s == 0) and (W / s == 0), 'Indivisible gain'
+        s = self.gain
+        x = x.view(b, c, h // s, s, w // s, s)  # x(1,64,40,2,40,2)
+        x = x.permute(0, 3, 5, 1, 2, 4).contiguous()  # x(1,2,2,64,40,40)
+        return x.view(b, c * s * s, h // s, w // s)  # x(1,256,40,40)
+
+
+class Expand(nn.Module):
+    # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        b, c, h, w = x.size()  # assert C / s ** 2 == 0, 'Indivisible gain'
+        s = self.gain
+        x = x.view(b, s, s, c // s ** 2, h, w)  # x(1,2,2,16,80,80)
+        x = x.permute(0, 3, 4, 1, 5, 2).contiguous()  # x(1,16,80,2,80,2)
+        return x.view(b, c // s ** 2, h * s, w * s)  # x(1,16,160,160)
+
+
+class Concat(nn.Module):
+    # Concatenate a list of tensors along dimension
+    def __init__(self, dimension=1):
+        super().__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        return torch.cat(x, self.d)
+
+
+class AutoShape(nn.Module):
+    # YOLOv5 input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
+    conf = 0.25  # NMS confidence threshold
+    iou = 0.45  # NMS IoU threshold
+    classes = None  # (optional list) filter by class
+    max_det = 1000  # maximum number of detections per image
+
+    def __init__(self, model):
+        super().__init__()
+        self.model = model.eval()
+
+    def autoshape(self):
+        LOGGER.info('AutoShape already enabled, skipping... ')  # model already converted to model.autoshape()
+        return self
+
+    @torch.no_grad()
+    def forward(self, imgs, size=640, augment=False, profile=False):
+        # Inference from various sources. For height=640, width=1280, RGB images example inputs are:
+        #   file:       imgs = 'data/images/zidane.jpg'  # str or PosixPath
+        #   URI:             = 'https://ultralytics.com/images/zidane.jpg'
+        #   OpenCV:          = cv2.imread('image.jpg')[:,:,::-1]  # HWC BGR to RGB x(640,1280,3)
+        #   PIL:             = Image.open('image.jpg') or ImageGrab.grab()  # HWC x(640,1280,3)
+        #   numpy:           = np.zeros((640,1280,3))  # HWC
+        #   torch:           = torch.zeros(16,3,320,640)  # BCHW (scaled to size=640, 0-1 values)
+        #   multiple:        = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...]  # list of images
+
+        t = [time_sync()]
+        p = next(self.model.parameters())  # for device and type
+        if isinstance(imgs, torch.Tensor):  # torch
+            with amp.autocast(enabled=p.device.type != 'cpu'):
+                return self.model(imgs.to(p.device).type_as(p), augment, profile)  # inference
+
+        # Pre-process
+        n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs])  # number of images, list of images
+        shape0, shape1, files = [], [], []  # image and inference shapes, filenames
+        for i, im in enumerate(imgs):
+            f = f'image{i}'  # filename
+            if isinstance(im, (str, Path)):  # filename or uri
+                im, f = Image.open(requests.get(im, stream=True).raw if str(im).startswith('http') else im), im
+                im = np.asarray(exif_transpose(im))
+            elif isinstance(im, Image.Image):  # PIL Image
+                im, f = np.asarray(exif_transpose(im)), getattr(im, 'filename', f) or f
+            files.append(Path(f).with_suffix('.jpg').name)
+            if im.shape[0] < 5:  # image in CHW
+                im = im.transpose((1, 2, 0))  # reverse dataloader .transpose(2, 0, 1)
+            im = im[..., :3] if im.ndim == 3 else np.tile(im[..., None], 3)  # enforce 3ch input
+            s = im.shape[:2]  # HWC
+            shape0.append(s)  # image shape
+            g = (size / max(s))  # gain
+            shape1.append([y * g for y in s])
+            imgs[i] = im if im.data.contiguous else np.ascontiguousarray(im)  # update
+        shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)]  # inference shape
+        x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs]  # pad
+        x = np.stack(x, 0) if n > 1 else x[0][None]  # stack
+        x = np.ascontiguousarray(x.transpose((0, 3, 1, 2)))  # BHWC to BCHW
+        x = torch.from_numpy(x).to(p.device).type_as(p) / 255.  # uint8 to fp16/32
+        t.append(time_sync())
+
+        with amp.autocast(enabled=p.device.type != 'cpu'):
+            # Inference
+            y = self.model(x, augment, profile)[0]  # forward
+            t.append(time_sync())
+
+            # Post-process
+            y = non_max_suppression(y, self.conf, iou_thres=self.iou, classes=self.classes, max_det=self.max_det)  # NMS
+            for i in range(n):
+                scale_coords(shape1, y[i][:, :4], shape0[i])
+
+            t.append(time_sync())
+            return Detections(imgs, y, files, t, self.names, x.shape)
+
+
+class Detections:
+    # YOLOv5 detections class for inference results
+    def __init__(self, imgs, pred, files, times=None, names=None, shape=None):
+        super().__init__()
+        d = pred[0].device  # device
+        gn = [torch.tensor([*[im.shape[i] for i in [1, 0, 1, 0]], 1., 1.], device=d) for im in imgs]  # normalizations
+        self.imgs = imgs  # list of images as numpy arrays
+        self.pred = pred  # list of tensors pred[0] = (xyxy, conf, cls)
+        self.names = names  # class names
+        self.ascii = is_ascii(names)  # names are ascii (use PIL for UTF-8)
+        self.files = files  # image filenames
+        self.xyxy = pred  # xyxy pixels
+        self.xywh = [xyxy2xywh(x) for x in pred]  # xywh pixels
+        self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)]  # xyxy normalized
+        self.xywhn = [x / g for x, g in zip(self.xywh, gn)]  # xywh normalized
+        self.n = len(self.pred)  # number of images (batch size)
+        self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3))  # timestamps (ms)
+        self.s = shape  # inference BCHW shape
+
+    def display(self, pprint=False, show=False, save=False, crop=False, render=False, save_dir=Path('')):
+        for i, (im, pred) in enumerate(zip(self.imgs, self.pred)):
+            str = f'image {i + 1}/{len(self.pred)}: {im.shape[0]}x{im.shape[1]} '
+            if pred.shape[0]:
+                for c in pred[:, -1].unique():
+                    n = (pred[:, -1] == c).sum()  # detections per class
+                    str += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, "  # add to string
+                if show or save or render or crop:
+                    annotator = Annotator(im, pil=not self.ascii)
+                    for *box, conf, cls in reversed(pred):  # xyxy, confidence, class
+                        label = f'{self.names[int(cls)]} {conf:.2f}'
+                        if crop:
+                            save_one_box(box, im, file=save_dir / 'crops' / self.names[int(cls)] / self.files[i])
+                        else:  # all others
+                            annotator.box_label(box, label, color=colors(cls))
+                    im = annotator.im
+            else:
+                str += '(no detections)'
+
+            im = Image.fromarray(im.astype(np.uint8)) if isinstance(im, np.ndarray) else im  # from np
+            if pprint:
+                LOGGER.info(str.rstrip(', '))
+            if show:
+                im.show(self.files[i])  # show
+            if save:
+                f = self.files[i]
+                im.save(save_dir / f)  # save
+                if i == self.n - 1:
+                    LOGGER.info(f"Saved {self.n} image{'s' * (self.n > 1)} to {colorstr('bold', save_dir)}")
+            if render:
+                self.imgs[i] = np.asarray(im)
+
+    def print(self):
+        self.display(pprint=True)  # print results
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' %
+                    self.t)
+
+    def show(self):
+        self.display(show=True)  # show results
+
+    def save(self, save_dir='runs/detect/exp'):
+        save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True)  # increment save_dir
+        self.display(save=True, save_dir=save_dir)  # save results
+
+    def crop(self, save_dir='runs/detect/exp'):
+        save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True)  # increment save_dir
+        self.display(crop=True, save_dir=save_dir)  # crop results
+        LOGGER.info(f'Saved results to {save_dir}\n')
+
+    def render(self):
+        self.display(render=True)  # render results
+        return self.imgs
+
+    def pandas(self):
+        # return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0])
+        new = copy(self)  # return copy
+        ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name'  # xyxy columns
+        cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name'  # xywh columns
+        for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
+            a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)]  # update
+            setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
+        return new
+
+    def tolist(self):
+        # return a list of Detections objects, i.e. 'for result in results.tolist():'
+        x = [Detections([self.imgs[i]], [self.pred[i]], self.names, self.s) for i in range(self.n)]
+        for d in x:
+            for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']:
+                setattr(d, k, getattr(d, k)[0])  # pop out of list
+        return x
+
+    def __len__(self):
+        return self.n
+
+
+class Classify(nn.Module):
+    # Classification head, i.e. x(b,c1,20,20) to x(b,c2)
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.aap = nn.AdaptiveAvgPool2d(1)  # to x(b,c1,1,1)
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g)  # to x(b,c2,1,1)
+        self.flat = nn.Flatten()
+
+    def forward(self, x):
+        z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1)  # cat if list
+        return self.flat(self.conv(z))  # flatten to x(b,c2)

models/experimental.py

@@ -0,0 +1,115 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Experimental modules
+"""
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from models.common import Conv
+from utils.downloads import attempt_download
+
+
+class CrossConv(nn.Module):
+    # Cross Convolution Downsample
+    def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
+        # ch_in, ch_out, kernel, stride, groups, expansion, shortcut
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, (1, k), (1, s))
+        self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class Sum(nn.Module):
+    # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
+    def __init__(self, n, weight=False):  # n: number of inputs
+        super().__init__()
+        self.weight = weight  # apply weights boolean
+        self.iter = range(n - 1)  # iter object
+        if weight:
+            self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True)  # layer weights
+
+    def forward(self, x):
+        y = x[0]  # no weight
+        if self.weight:
+            w = torch.sigmoid(self.w) * 2
+            for i in self.iter:
+                y = y + x[i + 1] * w[i]
+        else:
+            for i in self.iter:
+                y = y + x[i + 1]
+        return y
+
+
+class MixConv2d(nn.Module):
+    # Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595
+    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
+        super().__init__()
+        groups = len(k)
+        if equal_ch:  # equal c_ per group
+            i = torch.linspace(0, groups - 1E-6, c2).floor()  # c2 indices
+            c_ = [(i == g).sum() for g in range(groups)]  # intermediate channels
+        else:  # equal weight.numel() per group
+            b = [c2] + [0] * groups
+            a = np.eye(groups + 1, groups, k=-1)
+            a -= np.roll(a, 1, axis=1)
+            a *= np.array(k) ** 2
+            a[0] = 1
+            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b
+
+        self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
+        self.bn = nn.BatchNorm2d(c2)
+        self.act = nn.LeakyReLU(0.1, inplace=True)
+
+    def forward(self, x):
+        return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
+
+
+class Ensemble(nn.ModuleList):
+    # Ensemble of models
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, augment=False, profile=False, visualize=False):
+        y = []
+        for module in self:
+            y.append(module(x, augment, profile, visualize)[0])
+        # y = torch.stack(y).max(0)[0]  # max ensemble
+        # y = torch.stack(y).mean(0)  # mean ensemble
+        y = torch.cat(y, 1)  # nms ensemble
+        return y, None  # inference, train output
+
+
+def attempt_load(weights, map_location=None, inplace=True, fuse=True):
+    from models.yolo import Detect, Model
+
+    # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
+    model = Ensemble()
+    for w in weights if isinstance(weights, list) else [weights]:
+        ckpt = torch.load(attempt_download(w), map_location=map_location)  # load
+        if fuse:
+            model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval())  # FP32 model
+        else:
+            model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().eval())  # without layer fuse
+
+
+    # Compatibility updates
+    for m in model.modules():
+        if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:
+            m.inplace = inplace  # pytorch 1.7.0 compatibility
+        elif type(m) is Conv:
+            m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
+
+    if len(model) == 1:
+        return model[-1]  # return model
+    else:
+        print(f'Ensemble created with {weights}\n')
+        for k in ['names']:
+            setattr(model, k, getattr(model[-1], k))
+        model.stride = model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride  # max stride
+        return model  # return ensemble

models/yolo.py

@@ -0,0 +1,327 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+YOLO-specific modules
+
+Usage:
+    $ python path/to/models/yolo.py --cfg yolov5s.yaml
+"""
+
+import argparse
+import sys
+from copy import deepcopy
+from pathlib import Path
+
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[1].as_posix())  # add yolov5/ to path
+
+from models.common import *
+from models.experimental import *
+from utils.autoanchor import check_anchor_order
+from utils.general import make_divisible, check_file, set_logging
+from utils.plots import feature_visualization
+from utils.torch_utils import time_sync, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \
+    select_device, copy_attr
+
+try:
+    import thop  # for FLOPs computation
+except ImportError:
+    thop = None
+
+LOGGER = logging.getLogger(__name__)
+
+
+class Detect(nn.Module):
+    stride = None  # strides computed during build
+    onnx_dynamic = False  # ONNX export parameter
+
+    def __init__(self, nc=80, anchors=(), ch=(), inplace=True, num_coords=0):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
+        self.register_buffer('anchors', a)  # shape(nl,na,2)
+        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
+        self.inplace = inplace  # use in-place ops (e.g. slice assignment)
+        self.num_coords = num_coords
+
+    def forward(self, x):
+        z = []  # inference output
+        for i in range(self.nl):
+            x[i] = self.m[i](x[i])  # conv
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4] or self.onnx_dynamic:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+
+                y = x[i].sigmoid()
+                if self.inplace:
+                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+
+                    if hasattr(self, 'num_coords') and self.num_coords:
+                        y[..., -self.num_coords:] = y[..., -self.num_coords:] * 4. - 2.
+                        y[..., -self.num_coords:] *= self.anchor_grid[i].repeat((1, 1, 1, 1, self.num_coords // 2))
+                        y[..., -self.num_coords:] += (self.grid[i] * self.stride[i]).repeat((1, 1, 1, 1, self.num_coords // 2))
+
+                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
+                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2)  # wh
+                    y = torch.cat((xy, wh, y[..., 4:]), -1)
+                z.append(y.view(bs, -1, self.no))
+                # z.append(y)
+
+        return x if self.training else (torch.cat(z, 1), x)
+        # return x if self.training else (z, x)
+
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+
+
+class Model(nn.Module):
+    def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None, num_coords=0, autobalance=False):  # model, input channels, number of classes
+        super().__init__()
+        if isinstance(cfg, dict):
+            self.yaml = cfg  # model dict
+        else:  # is *.yaml
+            import yaml  # for torch hub
+            self.yaml_file = Path(cfg).name
+            with open(cfg) as f:
+                self.yaml = yaml.safe_load(f)  # model dict
+
+        # Define model
+        ch = self.yaml['ch'] = self.yaml.get('ch', ch)  # input channels
+        if nc + num_coords and nc + num_coords != self.yaml['nc']:
+            LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc + num_coords}")
+            self.yaml['nc'] = nc + num_coords  # override yaml value
+        if anchors:
+            LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}')
+            self.yaml['anchors'] = round(anchors)  # override yaml value
+        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
+        self.names = [str(i) for i in range(self.yaml['nc'])]  # default names
+        self.inplace = self.yaml.get('inplace', True)
+        self.num_coords = num_coords
+        if autobalance:
+            self.loss_coeffs = nn.Parameter(torch.zeros(2))
+        # LOGGER.info([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))])
+
+        # Build strides, anchors
+        m = self.model[-1]  # Detect()
+        if isinstance(m, Detect):
+            s = 256  # 2x min stride
+            m.inplace = self.inplace
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
+            m.anchors /= m.stride.view(-1, 1, 1)
+            check_anchor_order(m)
+            self.stride = m.stride
+            m.num_coords = self.num_coords
+            m.nc = nc
+            self._initialize_biases()  # only run once
+            # LOGGER.info('Strides: %s' % m.stride.tolist())
+
+        # Init weights, biases
+        initialize_weights(self)
+        self.info()
+        LOGGER.info('')
+
+    def forward(self, x, augment=False, profile=False, visualize=False, kp_flip=None,
+                scales=[0.5, 1, 2], flips=[None, 3, None]):
+        if augment:
+            return self.forward_augment(x, kp_flip, s=scales, f=flips)  # augmented inference, None
+        return self.forward_once(x, profile, visualize)  # single-scale inference, train
+
+    def forward_augment(self, x, kp_flip, s=[0.5, 1, 2], f=[None, 3, None]):
+        img_size = x.shape[-2:]  # height, width
+        # s = [1, 0.83, 0.67]  # scales
+        # f = [None, 3, None]  # flips (2-ud, 3-lr)
+        y = []  # outputs
+        train_out = None
+        for si, fi in zip(s, f):
+            xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
+            yi, train_out_i = self.forward_once(xi)  # forward
+            if si == 1 and fi is None:
+                train_out = train_out_i
+            # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
+            yi = self._descale_pred(yi, fi, si, img_size, kp_flip)
+            y.append(yi)
+        return torch.cat(y, 1), train_out  # augmented inference, train
+
+    def forward_once(self, x, profile=False, visualize=False):
+        y, dt = [], []  # outputs
+        for m in self.model:
+            if m.f != -1:  # if not from previous layer
+                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
+
+            if profile:
+                c = isinstance(m, Detect)  # copy input as inplace fix
+                o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0  # FLOPs
+                t = time_sync()
+                for _ in range(10):
+                    m(x.copy() if c else x)
+                dt.append((time_sync() - t) * 100)
+                if m == self.model[0]:
+                    LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s}  {'module'}")
+                LOGGER.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f}  {m.type}')
+
+            x = m(x)  # run
+            y.append(x if m.i in self.save else None)  # save output
+
+            if visualize:
+                feature_visualization(x, m.type, m.i, save_dir=visualize)
+
+        if profile:
+            LOGGER.info('%.1fms total' % sum(dt))
+        return x
+
+    def _descale_pred(self, p, flips, scale, img_size, kp_flip):
+        # de-scale predictions following augmented inference (inverse operation)
+        if self.inplace:
+            p[..., :4] /= scale  # de-scale bbox
+            if kp_flip:
+                p[..., -self.num_coords:] /= scale  # de-scale kp
+            if flips == 2:
+                p[..., 1] = img_size[0] - p[..., 1]  # de-flip ud
+            elif flips == 3:
+                p[..., 0] = img_size[1] - p[..., 0]  # de-flip lr
+                if kp_flip:
+                    p[..., 6:6 + self.nc - 1] = p[..., 6:6 + self.nc - 1][..., kp_flip]  # de-flip bbox conf
+                    p[..., -self.num_coords::2] = img_size[1] - p[..., -self.num_coords::2]  # de-flip kp x
+                    p[..., -self.num_coords::2] = p[..., -self.num_coords::2][..., kp_flip]  # swap lr kp (x)
+                    p[..., -self.num_coords + 1::2] = p[..., -self.num_coords + 1::2][..., kp_flip]  # swap lr kp (y)
+
+        else:
+            x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale  # de-scale
+            if flips == 2:
+                y = img_size[0] - y  # de-flip ud
+            elif flips == 3:
+                x = img_size[1] - x  # de-flip lr
+            p = torch.cat((x, y, wh, p[..., 4:]), -1)
+        return p
+
+    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
+        # https://arxiv.org/abs/1708.02002 section 3.3
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
+        m = self.model[-1]  # Detect() module
+        for mi, s in zip(m.m, m.stride):  # from
+            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b.data[:, 5:5+m.nc] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+
+    def _print_biases(self):
+        m = self.model[-1]  # Detect() module
+        for mi in m.m:  # from
+            b = mi.bias.detach().view(m.na, -1).T  # conv.bias(255) to (3,85)
+            LOGGER.info(
+                ('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean()))
+
+    # def _print_weights(self):
+    #     for m in self.model.modules():
+    #         if type(m) is Bottleneck:
+    #             LOGGER.info('%10.3g' % (m.w.detach().sigmoid() * 2))  # shortcut weights
+
+    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
+        LOGGER.info('Fusing layers... ')
+        for m in self.model.modules():
+            if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'):
+                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
+                delattr(m, 'bn')  # remove batchnorm
+                m.forward = m.forward_fuse  # update forward
+        self.info()
+        return self
+
+    def autoshape(self):  # add AutoShape module
+        LOGGER.info('Adding AutoShape... ')
+        m = AutoShape(self)  # wrap model
+        copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=())  # copy attributes
+        return m
+
+    def info(self, verbose=False, img_size=640):  # print model information
+        model_info(self, verbose, img_size)
+
+
+def parse_model(d, ch):  # model_dict, input_channels(3)
+    LOGGER.info('\n%3s%18s%3s%10s  %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))
+    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
+    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
+    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
+
+    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
+    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
+        m = eval(m) if isinstance(m, str) else m  # eval strings
+        for j, a in enumerate(args):
+            try:
+                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
+            except:
+                pass
+
+        n = n_ = max(round(n * gd), 1) if n > 1 else n  # depth gain
+        if m in [Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,
+                 BottleneckCSP, C3, C3TR, C3SPP, C3Ghost]:
+            c1, c2 = ch[f], args[0]
+            if c2 != no:  # if not output
+                c2 = make_divisible(c2 * gw, 8)
+
+            args = [c1, c2, *args[1:]]
+            if m in [BottleneckCSP, C3, C3TR, C3Ghost]:
+                args.insert(2, n)  # number of repeats
+                n = 1
+        elif m is nn.BatchNorm2d:
+            args = [ch[f]]
+        elif m is Concat:
+            c2 = sum([ch[x] for x in f])
+        elif m is Detect:
+            args.append([ch[x] for x in f])
+            if isinstance(args[1], int):  # number of anchors
+                args[1] = [list(range(args[1] * 2))] * len(f)
+        elif m is Contract:
+            c2 = ch[f] * args[0] ** 2
+        elif m is Expand:
+            c2 = ch[f] // args[0] ** 2
+        else:
+            c2 = ch[f]
+
+        m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args)  # module
+        t = str(m)[8:-2].replace('__main__.', '')  # module type
+        np = sum([x.numel() for x in m_.parameters()])  # number params
+        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
+        LOGGER.info('%3s%18s%3s%10.0f  %-40s%-30s' % (i, f, n_, np, t, args))  # print
+        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
+        layers.append(m_)
+        if i == 0:
+            ch = []
+        ch.append(c2)
+    return nn.Sequential(*layers), sorted(save)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--profile', action='store_true', help='profile model speed')
+    opt = parser.parse_args()
+    opt.cfg = check_file(opt.cfg)  # check file
+    set_logging()
+    device = select_device(opt.device)
+
+    # Create model
+    model = Model(opt.cfg).to(device)
+    model.train()
+
+    # Profile
+    if opt.profile:
+        img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device)
+        y = model(img, profile=True)
+
+    # Tensorboard (not working https://github.com/ultralytics/yolov5/issues/2898)
+    # from torch.utils.tensorboard import SummaryWriter
+    # tb_writer = SummaryWriter('.')
+    # LOGGER.info("Run 'tensorboard --logdir=models' to view tensorboard at http://localhost:6006/")
+    # tb_writer.add_graph(torch.jit.trace(model, img, strict=False), [])  # add model graph

models/yolov5l6.yaml

@@ -0,0 +1,60 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+anchors:
+  - [19,27,  44,40,  38,94]  # P3/8
+  - [96,68,  86,152,  180,137]  # P4/16
+  - [140,301,  303,264,  238,542]  # P5/32
+  - [436,615,  739,380,  925,792]  # P6/64
+
+# YOLOv5 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, C3, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
+   [-1, 9, C3, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
+   [-1, 9, C3, [512]],
+   [-1, 1, Conv, [768, 3, 2]],  # 7-P5/32
+   [-1, 3, C3, [768]],
+   [-1, 1, Conv, [1024, 3, 2]],  # 9-P6/64
+   [-1, 1, SPP, [1024, [3, 5, 7]]],
+   [-1, 3, C3, [1024, False]],  # 11
+  ]
+
+# YOLOv5 head
+head:
+  [[-1, 1, Conv, [768, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 8], 1, Concat, [1]],  # cat backbone P5
+   [-1, 3, C3, [768, False]],  # 15
+
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, C3, [512, False]],  # 19
+
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+   [-1, 3, C3, [256, False]],  # 23 (P3/8-small)
+
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, 20], 1, Concat, [1]],  # cat head P4
+   [-1, 3, C3, [512, False]],  # 26 (P4/16-medium)
+
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, 16], 1, Concat, [1]],  # cat head P5
+   [-1, 3, C3, [768, False]],  # 29 (P5/32-large)
+
+   [-1, 1, Conv, [768, 3, 2]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P6
+   [-1, 3, C3, [1024, False]],  # 32 (P6/64-xlarge)
+
+   [[23, 26, 29, 32], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5, P6)
+  ]

models/yolov5m6.yaml

@@ -0,0 +1,60 @@
+   # YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 0.67  # model depth multiple
+width_multiple: 0.75  # layer channel multiple
+anchors:
+  - [19,27,  44,40,  38,94]  # P3/8
+  - [96,68,  86,152,  180,137]  # P4/16
+  - [140,301,  303,264,  238,542]  # P5/32
+  - [436,615,  739,380,  925,792]  # P6/64
+
+# YOLOv5 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, C3, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
+   [-1, 9, C3, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
+   [-1, 9, C3, [512]],
+   [-1, 1, Conv, [768, 3, 2]],  # 7-P5/32
+   [-1, 3, C3, [768]],
+   [-1, 1, Conv, [1024, 3, 2]],  # 9-P6/64
+   [-1, 1, SPP, [1024, [3, 5, 7]]],
+   [-1, 3, C3, [1024, False]],  # 11
+  ]
+
+# YOLOv5 head
+head:
+  [[-1, 1, Conv, [768, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 8], 1, Concat, [1]],  # cat backbone P5
+   [-1, 3, C3, [768, False]],  # 15
+
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, C3, [512, False]],  # 19
+
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+   [-1, 3, C3, [256, False]],  # 23 (P3/8-small)
+
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, 20], 1, Concat, [1]],  # cat head P4
+   [-1, 3, C3, [512, False]],  # 26 (P4/16-medium)
+
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, 16], 1, Concat, [1]],  # cat head P5
+   [-1, 3, C3, [768, False]],  # 29 (P5/32-large)
+
+   [-1, 1, Conv, [768, 3, 2]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P6
+   [-1, 3, C3, [1024, False]],  # 32 (P6/64-xlarge)
+
+   [[23, 26, 29, 32], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5, P6)
+  ]

models/yolov5s6.yaml

@@ -0,0 +1,60 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 0.33  # model depth multiple
+width_multiple: 0.50  # layer channel multiple
+anchors:
+  - [19,27,  44,40,  38,94]  # P3/8
+  - [96,68,  86,152,  180,137]  # P4/16
+  - [140,301,  303,264,  238,542]  # P5/32
+  - [436,615,  739,380,  925,792]  # P6/64
+
+# YOLOv5 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, C3, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
+   [-1, 9, C3, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
+   [-1, 9, C3, [512]],
+   [-1, 1, Conv, [768, 3, 2]],  # 7-P5/32
+   [-1, 3, C3, [768]],
+   [-1, 1, Conv, [1024, 3, 2]],  # 9-P6/64
+   [-1, 1, SPP, [1024, [3, 5, 7]]],
+   [-1, 3, C3, [1024, False]],  # 11
+  ]
+
+# YOLOv5 head
+head:
+  [[-1, 1, Conv, [768, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 8], 1, Concat, [1]],  # cat backbone P5
+   [-1, 3, C3, [768, False]],  # 15
+
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, C3, [512, False]],  # 19
+
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+   [-1, 3, C3, [256, False]],  # 23 (P3/8-small)
+
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, 20], 1, Concat, [1]],  # cat head P4
+   [-1, 3, C3, [512, False]],  # 26 (P4/16-medium)
+
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, 16], 1, Concat, [1]],  # cat head P5
+   [-1, 3, C3, [768, False]],  # 29 (P5/32-large)
+
+   [-1, 1, Conv, [768, 3, 2]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P6
+   [-1, 3, C3, [1024, False]],  # 32 (P6/64-xlarge)
+
+   [[23, 26, 29, 32], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5, P6)
+  ]

train.py

@@ -0,0 +1,601 @@
+import argparse
+import logging
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.cuda import amp
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import Adam, SGD, lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[0].as_posix())  # add yolov5/ to path
+
+import val
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.datasets import create_dataloader
+from utils.general import labels_to_class_weights, increment_path, labels_to_image_weights, init_seeds, \
+    strip_optimizer, get_latest_run, check_dataset, check_file, check_img_size, \
+    print_mutation, set_logging, one_cycle, colorstr, methods
+from utils.downloads import attempt_download
+from utils.loss import ComputeLoss
+from utils.plots import plot_labels, plot_evolve
+from utils.torch_utils import EarlyStopping, ModelEMA, de_parallel, intersect_dicts, select_device, \
+    torch_distributed_zero_first
+from utils.metrics import fitness
+from utils.loggers import Loggers
+from utils.callbacks import Callbacks
+
+LOGGER = logging.getLogger(__name__)
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+
+
+def train(hyp,  # path/to/hyp.yaml or hyp dictionary
+          opt,
+          device,
+          callbacks=Callbacks()
+          ):
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze, \
+        val_scales, val_flips = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze, opt.val_scales, opt.val_flips
+
+    val_flips = [None if f == -1 else f for f in val_flips]
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    w.mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp) as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+
+    # Save run settings
+    with open(save_dir / 'hyp.yaml', 'w') as f:
+        yaml.safe_dump(hyp, f, sort_keys=False)
+    with open(save_dir / 'opt.yaml', 'w') as f:
+        yaml.safe_dump(vars(opt), f, sort_keys=False)
+    data_dict = None
+
+    # Loggers
+    if RANK in [-1, 0]:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+        if loggers.wandb:
+            data_dict = loggers.wandb.data_dict
+            if resume:
+                weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+    # Config
+    plots = not evolve  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(1 + RANK)
+    with torch_distributed_zero_first(RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = ['item'] if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    assert len(names) == nc, f'{len(names)} names found for nc={nc} dataset in {data}'  # check
+    is_coco = data.endswith('coco.yaml') and nc == 80  # COCO dataset
+
+    labels_dir = data_dict.get('labels', 'labels')
+    kp_flip = data_dict.get('kp_flip')
+    kp_bbox = data_dict.get('kp_bbox')
+    num_coords = data_dict.get('num_coords', 0)
+
+    # Model
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location=device)  # load checkpoint
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors'), num_coords=num_coords).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors'), num_coords=num_coords).to(device)  # create
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in range(freeze)]  # layers to freeze
+    for k, v in model.named_parameters():
+        v.requires_grad = True  # train all layers
+        if any(x in k for x in freeze):
+            print(f'freezing {k}')
+            v.requires_grad = False
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    LOGGER.info(f"Scaled weight_decay = {hyp['weight_decay']}")
+
+    g0, g1, g2 = [], [], []  # optimizer parameter groups
+    for v in model.modules():
+        if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):  # bias
+            g2.append(v.bias)
+        if isinstance(v, nn.BatchNorm2d):  # weight (no decay)
+            g0.append(v.weight)
+        elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):  # weight (with decay)
+            g1.append(v.weight)
+
+    if opt.adam:
+        optimizer = Adam(g0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999))  # adjust beta1 to momentum
+    else:
+        optimizer = SGD(g0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
+
+    optimizer.add_param_group({'params': g1, 'weight_decay': hyp['weight_decay']})  # add g1 with weight_decay
+    optimizer.add_param_group({'params': g2})  # add g2 (biases)
+    # if opt.autobalance:
+    #     optimizer.add_param_group({'params': model.loss_coeffs})  # for autobalancing if used
+
+    LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups "
+                f"{len(g0)} weight, {len(g1)} weight (no decay), {len(g2)} bias")
+    del g0, g1, g2
+
+    # Scheduler
+    if opt.linear_lr:
+        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+    else:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in [-1, 0] else None
+
+    # Resume
+    start_epoch, best_fitness = 0, 0.0
+    if pretrained:
+        # Optimizer
+        if ckpt['optimizer'] is not None:
+            optimizer.load_state_dict(ckpt['optimizer'])
+            best_fitness = ckpt['best_fitness']
+
+        # EMA
+        if ema and ckpt.get('ema'):
+            ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
+            ema.updates = ckpt['updates']
+
+        # Epochs
+        start_epoch = ckpt['epoch'] + 1
+        if resume:
+            assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.'
+        if epochs < start_epoch:
+            LOGGER.info(f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs.")
+            epochs += ckpt['epoch']  # finetune additional epochs
+
+        del ckpt, csd
+
+    # Image sizes
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    nl = model.model[-1].nl  # number of detection layers (used for scaling hyp['obj'])
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        logging.warning('DP not recommended, instead use torch.distributed.run for best DDP Multi-GPU results.\n'
+                        'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path, labels_dir, imgsz, batch_size // WORLD_SIZE, gs, single_cls,
+                                              hyp=hyp, augment=True, cache=opt.cache, rect=opt.rect, rank=RANK,
+                                              workers=workers, image_weights=opt.image_weights, quad=opt.quad,
+                                              prefix=colorstr('train: '), kp_flip=kp_flip, kp_bbox=kp_bbox)
+    mlc = int(np.concatenate(dataset.labels, 0)[:, 0].max())  # max label class
+    nb = len(train_loader)  # number of batches
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in [-1, 0]:
+        val_loader = create_dataloader(val_path, labels_dir, imgsz, batch_size // WORLD_SIZE, gs, single_cls,
+                                       hyp=hyp, cache=None if noval else opt.cache, rect=False, rank=-1,
+                                       workers=workers, pad=0.5,
+                                       prefix=colorstr('val: '), kp_flip=kp_flip, kp_bbox=kp_bbox)[0]
+
+        if not resume:
+            # labels = np.concatenate(dataset.labels, 0)
+            # c = torch.tensor(labels[:, 0])  # classes
+            # cf = torch.bincount(c.long(), minlength=nc) + 1.  # frequency
+            # model._initialize_biases(cf.to(device))
+            # if plots:
+            #     plot_labels(labels, names, save_dir)
+
+            # Anchors
+            if not opt.noautoanchor:
+                check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.on_pretrain_routine_end()
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)
+
+    # Model parameters
+    hyp['box'] *= 3. / nl  # scale to layers
+    hyp['cls'] *= nc / 80. * 3. / nl  # scale to classes and layers
+    hyp['obj'] *= (imgsz / 640) ** 2 * 3. / nl  # scale to image size and layers
+    hyp['kp'] *= 3. / nl
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nw = max(round(hyp['warmup_epochs'] * nb), 1000)  # number of warmup iterations, max(3 epochs, 1k iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls, kp)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = amp.GradScaler(enabled=cuda)
+    stopper = EarlyStopping(patience=opt.patience)
+    compute_loss = ComputeLoss(model, autobalance=False, num_coords=num_coords)  # init loss class
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(4, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'kps', 'labels', 'img_size'))
+        if RANK in [-1, 0]:
+            pbar = tqdm(pbar, total=nb)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255.0  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with amp.autocast(enabled=cuda):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize
+            if ni - last_opt_step >= accumulate:
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in [-1, 0]:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%10s' * 2 + '%10.4g' * 6) % (
+                    f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.on_train_batch_end(ni, model, imgs, targets, paths, plots, opt.sync_bn)
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups[:3]]  # for loggers
+        scheduler.step()
+
+        if RANK in [-1, 0]:
+            # mAP
+            callbacks.on_train_epoch_end(epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = val.run(data_dict,
+                                           batch_size=batch_size // WORLD_SIZE,
+                                           imgsz=imgsz,
+                                           conf_thres=0.01,
+                                           model=ema.ema,
+                                           dataloader=val_loader,
+                                           compute_loss=compute_loss,
+                                           scales=val_scales,
+                                           flips=val_flips)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.on_fit_epoch_end(log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {'epoch': epoch,
+                        'best_fitness': best_fitness,
+                        'model': deepcopy(de_parallel(model)).half(),
+                        'ema': deepcopy(ema.ema).half(),
+                        'updates': ema.updates,
+                        'optimizer': optimizer.state_dict(),
+                        'wandb_id': loggers.wandb.wandb_run.id if loggers.wandb else None}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                del ckpt
+                callbacks.on_model_save(last, epoch, final_epoch, best_fitness, fi)
+
+            # Stop Single-GPU
+            if RANK == -1 and stopper(epoch=epoch, fitness=fi):
+                break
+
+            # Stop DDP TODO: known issues shttps://github.com/ultralytics/yolov5/pull/4576
+            # stop = stopper(epoch=epoch, fitness=fi)
+            # if RANK == 0:
+            #    dist.broadcast_object_list([stop], 0)  # broadcast 'stop' to all ranks
+
+        # Stop DPP
+        # with torch_distributed_zero_first(RANK):
+        # if stop:
+        #    break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in [-1, 0]:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        if not evolve:
+            # Strip optimizers
+            for f in last, best:
+                if f.exists():
+                    strip_optimizer(f)  # strip optimizers
+        callbacks.on_train_end(last, best, plots, epoch)
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default='yolov5s6.pt', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--data', type=str, default='data/coco-kp.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default='data/hyps/hyp.kp-p6.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=300)
+    parser.add_argument('--batch-size', type=int, default=8, help='total batch size for all GPUs')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=1280, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--adam', action='store_true', help='use torch.optim.Adam() optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers')
+    parser.add_argument('--project', default='runs/train', help='save to project/name')
+    parser.add_argument('--entity', default=None, help='W&B entity')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--linear-lr', action='store_true', help='linear LR')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--upload_dataset', action='store_true', help='Upload dataset as W&B artifact table')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval for W&B')
+    parser.add_argument('--save_period', type=int, default=-1, help='Log model after every "save_period" epoch')
+    parser.add_argument('--artifact_alias', type=str, default="latest", help='version of dataset artifact to be used')
+    parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')
+    parser.add_argument('--freeze', type=int, default=0, help='Number of layers to freeze. backbone=10, all=24')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--val-scales', type=float, nargs='+', default=[1])
+    parser.add_argument('--val-flips', type=int, nargs='+', default=[-1])
+    parser.add_argument('--autobalance', action='store_true', help='Learn keypoint and object loss scaling')
+    opt = parser.parse_known_args()[0] if known else parser.parse_args()
+    return opt
+
+
+def main(opt):
+    # Checks
+    set_logging(RANK)
+    if RANK in [-1, 0]:
+        print(colorstr('train: ') + ', '.join(f'{k}={v}' for k, v in vars(opt).items()))
+        # check_git_status()
+        # check_requirements(requirements=FILE.parent / 'requirements.txt', exclude=['thop'])
+
+    # Resume
+    if opt.resume and not opt.evolve:  # resume an interrupted run
+        ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run()  # specified or most recent path
+        assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist'
+        with open(Path(ckpt).parent.parent / 'opt.yaml') as f:
+            opt = argparse.Namespace(**yaml.safe_load(f))  # replace
+        opt.cfg, opt.weights, opt.resume = '', ckpt, True  # reinstate
+        LOGGER.info(f'Resuming training from {ckpt}')
+    else:
+        opt.data, opt.cfg, opt.hyp = check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp)  # check files
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            opt.project = 'runs/evolve'
+            opt.exist_ok = opt.resume
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        from datetime import timedelta
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        assert opt.batch_size % WORLD_SIZE == 0, '--batch-size must be multiple of CUDA device count'
+        assert not opt.image_weights, '--image-weights argument is not compatible with DDP training'
+        assert not opt.evolve, '--evolve argument is not compatible with DDP training'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device)
+        if WORLD_SIZE > 1 and RANK == 0:
+            _ = [print('Destroying process group... ', end=''), dist.destroy_process_group(), print('Done.')]
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+                'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+                'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+                'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+                'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+                'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+                'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+                'box': (1, 0.02, 0.2),  # box loss gain
+                'cls': (1, 0.2, 4.0),  # cls loss gain
+                'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+                'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+                'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+                'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+                'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+                'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+                'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+                'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+                'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+                'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+                'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+                'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+                'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+                'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+                'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+                'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+                'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+                'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+                'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+                'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp) as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {save_dir}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device)
+
+            # Write mutation results
+            print_mutation(results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        print(f'Hyperparameter evolution finished\n'
+              f"Results saved to {colorstr('bold', save_dir)}\n"
+              f'Use best hyperparameters example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolov5m.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

utils/activations.py

@@ -0,0 +1,101 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Activation functions
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+# SiLU https://arxiv.org/pdf/1606.08415.pdf ----------------------------------------------------------------------------
+class SiLU(nn.Module):  # export-friendly version of nn.SiLU()
+    @staticmethod
+    def forward(x):
+        return x * torch.sigmoid(x)
+
+
+class Hardswish(nn.Module):  # export-friendly version of nn.Hardswish()
+    @staticmethod
+    def forward(x):
+        # return x * F.hardsigmoid(x)  # for torchscript and CoreML
+        return x * F.hardtanh(x + 3, 0., 6.) / 6.  # for torchscript, CoreML and ONNX
+
+
+# Mish https://github.com/digantamisra98/Mish --------------------------------------------------------------------------
+class Mish(nn.Module):
+    @staticmethod
+    def forward(x):
+        return x * F.softplus(x).tanh()
+
+
+class MemoryEfficientMish(nn.Module):
+    class F(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x):
+            ctx.save_for_backward(x)
+            return x.mul(torch.tanh(F.softplus(x)))  # x * tanh(ln(1 + exp(x)))
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            x = ctx.saved_tensors[0]
+            sx = torch.sigmoid(x)
+            fx = F.softplus(x).tanh()
+            return grad_output * (fx + x * sx * (1 - fx * fx))
+
+    def forward(self, x):
+        return self.F.apply(x)
+
+
+# FReLU https://arxiv.org/abs/2007.11824 -------------------------------------------------------------------------------
+class FReLU(nn.Module):
+    def __init__(self, c1, k=3):  # ch_in, kernel
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False)
+        self.bn = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        return torch.max(x, self.bn(self.conv(x)))
+
+
+# ACON https://arxiv.org/pdf/2009.04759.pdf ----------------------------------------------------------------------------
+class AconC(nn.Module):
+    r""" ACON activation (activate or not).
+    AconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is a learnable parameter
+    according to "Activate or Not: Learning Customized Activation" <https://arxiv.org/pdf/2009.04759.pdf>.
+    """
+
+    def __init__(self, c1):
+        super().__init__()
+        self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.beta = nn.Parameter(torch.ones(1, c1, 1, 1))
+
+    def forward(self, x):
+        dpx = (self.p1 - self.p2) * x
+        return dpx * torch.sigmoid(self.beta * dpx) + self.p2 * x
+
+
+class MetaAconC(nn.Module):
+    r""" ACON activation (activate or not).
+    MetaAconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is generated by a small network
+    according to "Activate or Not: Learning Customized Activation" <https://arxiv.org/pdf/2009.04759.pdf>.
+    """
+
+    def __init__(self, c1, k=1, s=1, r=16):  # ch_in, kernel, stride, r
+        super().__init__()
+        c2 = max(r, c1 // r)
+        self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.fc1 = nn.Conv2d(c1, c2, k, s, bias=True)
+        self.fc2 = nn.Conv2d(c2, c1, k, s, bias=True)
+        # self.bn1 = nn.BatchNorm2d(c2)
+        # self.bn2 = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        y = x.mean(dim=2, keepdims=True).mean(dim=3, keepdims=True)
+        # batch-size 1 bug/instabilities https://github.com/ultralytics/yolov5/issues/2891
+        # beta = torch.sigmoid(self.bn2(self.fc2(self.bn1(self.fc1(y)))))  # bug/unstable
+        beta = torch.sigmoid(self.fc2(self.fc1(y)))  # bug patch BN layers removed
+        dpx = (self.p1 - self.p2) * x
+        return dpx * torch.sigmoid(beta * dpx) + self.p2 * x

utils/augmentations.py

@@ -0,0 +1,321 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Image augmentation functions
+"""
+
+import logging
+import math
+import random
+
+import cv2
+import numpy as np
+
+from utils.general import colorstr, segment2box, resample_segments, check_version
+from utils.metrics import bbox_ioa
+
+
+class Albumentations:
+    # YOLOv5 Albumentations class (optional, only used if package is installed)
+    def __init__(self):
+        self.transform = None
+        try:
+            import albumentations as A
+            check_version(A.__version__, '1.0.3')  # version requirement
+
+            self.transform = A.Compose([
+                A.Blur(p=0.1),
+                A.MedianBlur(p=0.1),
+                A.ToGray(p=0.01)],
+                bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))
+
+            logging.info(colorstr('albumentations: ') + ', '.join(f'{x}' for x in self.transform.transforms if x.p))
+        except ImportError:  # package not installed, skip
+            pass
+        except Exception as e:
+            logging.info(colorstr('albumentations: ') + f'{e}')
+
+    def __call__(self, im, labels, p=1.0):
+        if self.transform and random.random() < p:
+            new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0])  # transformed
+            im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
+        return im, labels
+
+
+def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):
+    # HSV color-space augmentation
+    if hgain or sgain or vgain:
+        r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
+        hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))
+        dtype = im.dtype  # uint8
+
+        x = np.arange(0, 256, dtype=r.dtype)
+        lut_hue = ((x * r[0]) % 180).astype(dtype)
+        lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
+        lut_val = np.clip(x * r[2], 0, 255).astype(dtype)
+
+        im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
+        cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im)  # no return needed
+
+
+def hist_equalize(im, clahe=True, bgr=False):
+    # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-255
+    yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)
+    if clahe:
+        c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
+        yuv[:, :, 0] = c.apply(yuv[:, :, 0])
+    else:
+        yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0])  # equalize Y channel histogram
+    return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB)  # convert YUV image to RGB
+
+
+def replicate(im, labels):
+    # Replicate labels
+    h, w = im.shape[:2]
+    boxes = labels[:, 1:].astype(int)
+    x1, y1, x2, y2 = boxes.T
+    s = ((x2 - x1) + (y2 - y1)) / 2  # side length (pixels)
+    for i in s.argsort()[:round(s.size * 0.5)]:  # smallest indices
+        x1b, y1b, x2b, y2b = boxes[i]
+        bh, bw = y2b - y1b, x2b - x1b
+        yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw))  # offset x, y
+        x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]
+        im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b]  # im4[ymin:ymax, xmin:xmax]
+        labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)
+
+    return im, labels
+
+
+def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
+    # Resize and pad image while meeting stride-multiple constraints
+    shape = im.shape[:2]  # current shape [height, width]
+    if isinstance(new_shape, int):
+        new_shape = (new_shape, new_shape)
+
+    # Scale ratio (new / old)
+    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+    if not scaleup:  # only scale down, do not scale up (for better val mAP)
+        r = min(r, 1.0)
+
+    # Compute padding
+    ratio = r, r  # width, height ratios
+    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
+    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
+    if auto:  # minimum rectangle
+        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
+    elif scaleFill:  # stretch
+        dw, dh = 0.0, 0.0
+        new_unpad = (new_shape[1], new_shape[0])
+        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios
+
+    dw /= 2  # divide padding into 2 sides
+    dh /= 2
+
+    if shape[::-1] != new_unpad:  # resize
+        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
+    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
+    return im, ratio, (dw, dh)
+
+
+def random_perspective(im, targets=(), segments=(), degrees=10, translate=.1, scale=.1, shear=10, perspective=0.0,
+                       border=(0, 0), kp_bbox=None):
+    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
+    # targets = [cls, xyxy]
+
+    height = im.shape[0] + border[0] * 2  # shape(h,w,c)
+    width = im.shape[1] + border[1] * 2
+
+    # Center
+    C = np.eye(3)
+    C[0, 2] = -im.shape[1] / 2  # x translation (pixels)
+    C[1, 2] = -im.shape[0] / 2  # y translation (pixels)
+
+    # Perspective
+    P = np.eye(3)
+    P[2, 0] = random.uniform(-perspective, perspective)  # x perspective (about y)
+    P[2, 1] = random.uniform(-perspective, perspective)  # y perspective (about x)
+
+    # Rotation and Scale
+    R = np.eye(3)
+    a = random.uniform(-degrees, degrees)
+    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
+    s = random.uniform(1 - scale, 1 + scale)
+    # s = 2 ** random.uniform(-scale, scale)
+    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
+
+    # Shear
+    S = np.eye(3)
+    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
+    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)
+
+    # Translation
+    T = np.eye(3)
+    T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width  # x translation (pixels)
+    T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height  # y translation (pixels)
+
+    # Combined rotation matrix
+    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
+    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
+        if perspective:
+            im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))
+        else:  # affine
+            im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
+
+    # Visualize
+    # import matplotlib.pyplot as plt
+    # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
+    # ax[0].imshow(im[:, :, ::-1])  # base
+    # ax[1].imshow(im2[:, :, ::-1])  # warped
+
+    # Transform label coordinates
+    n = len(targets)
+    if n:
+        use_segments = any(x.any() for x in segments)
+        new = np.zeros((n, 4))
+        if use_segments:  # warp segments
+            segments = resample_segments(segments)  # upsample
+            for i, segment in enumerate(segments):
+                xy = np.ones((len(segment), 3))
+                xy[:, :2] = segment
+                xy = xy @ M.T  # transform
+                xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]  # perspective rescale or affine
+
+                # clip
+                new[i] = segment2box(xy, width, height)
+
+        else:  # warp boxes
+            xy = np.ones((n * 4, 3))
+            xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
+            xy = xy @ M.T  # transform
+            xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine
+
+            # create new boxes
+            x = xy[:, [0, 2, 4, 6]]
+            y = xy[:, [1, 3, 5, 7]]
+            new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
+
+            # clip
+            new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)
+            new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)
+
+        # filter candidates
+        i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)
+        targets = targets[i]
+        targets[:, 1:5] = new[i]
+
+        n = len(targets)
+        if n and targets.shape[1] > 5:
+            # warp keypoints in person object
+            person_mask = targets[:, 0] == 0
+            person_targets = targets[person_mask]
+            if len(person_targets) > 0:
+                xy = person_targets[:, 5:].reshape(-1, 3)
+                vis = xy[:, 2:].copy()
+                xy[:, 2:] = 1
+                xy = xy @ M.T  # transform
+                xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2])  # perspective rescale or affine
+                out_mask = (
+                        (xy[:, 0] < 0) |
+                        (xy[:, 1] < 0) |
+                        (xy[:, 0] > width) |
+                        (xy[:, 1] > height)
+                )
+                vis[out_mask] = 0
+                keypoints = np.concatenate((xy, vis), axis=-1)
+                targets[person_mask, 5:] = keypoints.reshape(person_targets.shape[0], -1)
+
+        # resize keypoint bbox sizes back to original
+        if n and kp_bbox is not None:
+            for i in range(int(targets[:, 0].max()) + 1):
+                if i > 0:
+                    if isinstance(kp_bbox, list):
+                        kp_bbox_i = kp_bbox[i - 1]
+                    else:
+                        kp_bbox_i = kp_bbox
+
+                    kp_mask = targets[:, 0] == i
+                    kp_targets = targets[kp_mask]
+
+                    xc = kp_targets[:, [1, 3]].mean(axis=-1)
+                    yc = kp_targets[:, [2, 4]].mean(axis=-1)
+
+                    kp_targets[:, 1] = xc - (kp_bbox_i * width) / 2
+                    kp_targets[:, 2] = yc - (kp_bbox_i * height) / 2
+                    kp_targets[:, 3] = xc + (kp_bbox_i * width) / 2
+                    kp_targets[:, 4] = yc + (kp_bbox_i * height) / 2
+
+                    targets[kp_mask] = kp_targets
+
+                # clip
+                targets[:, [1, 3]] = targets[:, [1, 3]].clip(0, width)
+                targets[:, [2, 4]] = targets[:, [2, 4]].clip(0, height)
+
+    return im, targets
+
+
+def copy_paste(im, labels, segments, p=0.5):
+    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
+    n = len(segments)
+    if p and n:
+        h, w, c = im.shape  # height, width, channels
+        im_new = np.zeros(im.shape, np.uint8)
+        for j in random.sample(range(n), k=round(p * n)):
+            l, s = labels[j], segments[j]
+            box = w - l[3], l[2], w - l[1], l[4]
+            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
+            if (ioa < 0.30).all():  # allow 30% obscuration of existing labels
+                labels = np.concatenate((labels, [[l[0], *box]]), 0)
+                segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))
+                cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
+
+        result = cv2.bitwise_and(src1=im, src2=im_new)
+        result = cv2.flip(result, 1)  # augment segments (flip left-right)
+        i = result > 0  # pixels to replace
+        # i[:, :] = result.max(2).reshape(h, w, 1)  # act over ch
+        im[i] = result[i]  # cv2.imwrite('debug.jpg', im)  # debug
+
+    return im, labels, segments
+
+
+def cutout(im, labels, p=0.5):
+    # Applies image cutout augmentation https://arxiv.org/abs/1708.04552
+    if random.random() < p:
+        h, w = im.shape[:2]
+        scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16  # image size fraction
+        for s in scales:
+            mask_h = random.randint(1, int(h * s))  # create random masks
+            mask_w = random.randint(1, int(w * s))
+
+            # box
+            xmin = max(0, random.randint(0, w) - mask_w // 2)
+            ymin = max(0, random.randint(0, h) - mask_h // 2)
+            xmax = min(w, xmin + mask_w)
+            ymax = min(h, ymin + mask_h)
+
+            # apply random color mask
+            im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]
+
+            # return unobscured labels
+            if len(labels) and s > 0.03:
+                box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
+                ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
+                labels = labels[ioa < 0.60]  # remove >60% obscured labels
+
+    return labels
+
+
+def mixup(im, labels, im2, labels2):
+    # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
+    r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
+    im = (im * r + im2 * (1 - r)).astype(np.uint8)
+    labels = np.concatenate((labels, labels2), 0)
+    return im, labels
+
+
+def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.1, eps=1e-16):  # box1(4,n), box2(4,n)
+    # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
+    w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
+    w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
+    ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
+    return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates

utils/autoanchor.py

@@ -0,0 +1,164 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Auto-anchor utils
+"""
+
+import random
+
+import numpy as np
+import torch
+import yaml
+from tqdm import tqdm
+
+from utils.general import colorstr
+
+
+def check_anchor_order(m):
+    # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
+    a = m.anchor_grid.prod(-1).view(-1)  # anchor area
+    da = a[-1] - a[0]  # delta a
+    ds = m.stride[-1] - m.stride[0]  # delta s
+    if da.sign() != ds.sign():  # same order
+        print('Reversing anchor order')
+        m.anchors[:] = m.anchors.flip(0)
+        m.anchor_grid[:] = m.anchor_grid.flip(0)
+
+
+def check_anchors(dataset, model, thr=4.0, imgsz=640):
+    # Check anchor fit to data, recompute if necessary
+    prefix = colorstr('autoanchor: ')
+    print(f'\n{prefix}Analyzing anchors... ', end='')
+    m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]  # Detect()
+    shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh
+
+    def metric(k):  # compute metric
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
+        best = x.max(1)[0]  # best_x
+        aat = (x > 1. / thr).float().sum(1).mean()  # anchors above threshold
+        bpr = (best > 1. / thr).float().mean()  # best possible recall
+        return bpr, aat
+
+    anchors = m.anchor_grid.clone().cpu().view(-1, 2)  # current anchors
+    bpr, aat = metric(anchors)
+    print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='')
+    if bpr < 0.98:  # threshold to recompute
+        print('. Attempting to improve anchors, please wait...')
+        na = m.anchor_grid.numel() // 2  # number of anchors
+        try:
+            anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
+        except Exception as e:
+            print(f'{prefix}ERROR: {e}')
+        new_bpr = metric(anchors)[0]
+        if new_bpr > bpr:  # replace anchors
+            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
+            m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid)  # for inference
+            m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1)  # loss
+            check_anchor_order(m)
+            print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.')
+        else:
+            print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.')
+    print('')  # newline
+
+
+def kmean_anchors(dataset='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
+    """ Creates kmeans-evolved anchors from training dataset
+
+        Arguments:
+            dataset: path to data.yaml, or a loaded dataset
+            n: number of anchors
+            img_size: image size used for training
+            thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
+            gen: generations to evolve anchors using genetic algorithm
+            verbose: print all results
+
+        Return:
+            k: kmeans evolved anchors
+
+        Usage:
+            from utils.autoanchor import *; _ = kmean_anchors()
+    """
+    from scipy.cluster.vq import kmeans
+
+    thr = 1. / thr
+    prefix = colorstr('autoanchor: ')
+
+    def metric(k, wh):  # compute metrics
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
+        # x = wh_iou(wh, torch.tensor(k))  # iou metric
+        return x, x.max(1)[0]  # x, best_x
+
+    def anchor_fitness(k):  # mutation fitness
+        _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
+        return (best * (best > thr).float()).mean()  # fitness
+
+    def print_results(k):
+        k = k[np.argsort(k.prod(1))]  # sort small to large
+        x, best = metric(k, wh0)
+        bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
+        print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr')
+        print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, '
+              f'past_thr={x[x > thr].mean():.3f}-mean: ', end='')
+        for i, x in enumerate(k):
+            print('%i,%i' % (round(x[0]), round(x[1])), end=',  ' if i < len(k) - 1 else '\n')  # use in *.cfg
+        return k
+
+    if isinstance(dataset, str):  # *.yaml file
+        with open(dataset, errors='ignore') as f:
+            data_dict = yaml.safe_load(f)  # model dict
+        from utils.datasets import LoadImagesAndLabels
+        dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
+
+    # Get label wh
+    shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
+
+    # Filter
+    i = (wh0 < 3.0).any(1).sum()
+    if i:
+        print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.')
+    wh = wh0[(wh0 >= 2.0).any(1)]  # filter > 2 pixels
+    # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1
+
+    # Kmeans calculation
+    print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...')
+    s = wh.std(0)  # sigmas for whitening
+    k, dist = kmeans(wh / s, n, iter=30)  # points, mean distance
+    assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}')
+    k *= s
+    wh = torch.tensor(wh, dtype=torch.float32)  # filtered
+    wh0 = torch.tensor(wh0, dtype=torch.float32)  # unfiltered
+    k = print_results(k)
+
+    # Plot
+    # k, d = [None] * 20, [None] * 20
+    # for i in tqdm(range(1, 21)):
+    #     k[i-1], d[i-1] = kmeans(wh / s, i)  # points, mean distance
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
+    # ax = ax.ravel()
+    # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7))  # plot wh
+    # ax[0].hist(wh[wh[:, 0]<100, 0],400)
+    # ax[1].hist(wh[wh[:, 1]<100, 1],400)
+    # fig.savefig('wh.png', dpi=200)
+
+    # Evolve
+    npr = np.random
+    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
+    pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:')  # progress bar
+    for _ in pbar:
+        v = np.ones(sh)
+        while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
+            v = ((npr.random(sh) < mp) * random.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
+        kg = (k.copy() * v).clip(min=2.0)
+        fg = anchor_fitness(kg)
+        if fg > f:
+            f, k = fg, kg.copy()
+            pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
+            if verbose:
+                print_results(k)
+
+    return print_results(k)

utils/callbacks.py

@@ -0,0 +1,179 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Callback utils
+"""
+
+
+class Callbacks:
+    """"
+    Handles all registered callbacks for YOLOv5 Hooks
+    """
+
+    _callbacks = {
+        'on_pretrain_routine_start': [],
+        'on_pretrain_routine_end': [],
+
+        'on_train_start': [],
+        'on_train_epoch_start': [],
+        'on_train_batch_start': [],
+        'optimizer_step': [],
+        'on_before_zero_grad': [],
+        'on_train_batch_end': [],
+        'on_train_epoch_end': [],
+
+        'on_val_start': [],
+        'on_val_batch_start': [],
+        'on_val_image_end': [],
+        'on_val_batch_end': [],
+        'on_val_end': [],
+
+        'on_fit_epoch_end': [],  # fit = train + val
+        'on_model_save': [],
+        'on_train_end': [],
+
+        'teardown': [],
+    }
+
+    def __init__(self):
+        return
+
+    def register_action(self, hook, name='', callback=None):
+        """
+        Register a new action to a callback hook
+
+        Args:
+            hook        The callback hook name to register the action to
+            name        The name of the action
+            callback    The callback to fire
+        """
+        assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}"
+        assert callable(callback), f"callback '{callback}' is not callable"
+        self._callbacks[hook].append({'name': name, 'callback': callback})
+
+    def get_registered_actions(self, hook=None):
+        """"
+        Returns all the registered actions by callback hook
+
+        Args:
+            hook The name of the hook to check, defaults to all
+        """
+        if hook:
+            return self._callbacks[hook]
+        else:
+            return self._callbacks
+
+    def run_callbacks(self, hook, *args, **kwargs):
+        """
+        Loop through the registered actions and fire all callbacks
+        """
+        for logger in self._callbacks[hook]:
+            # print(f"Running callbacks.{logger['callback'].__name__}()")
+            logger['callback'](*args, **kwargs)
+
+    def on_pretrain_routine_start(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the start of each pretraining routine
+        """
+        self.run_callbacks('on_pretrain_routine_start', *args, **kwargs)
+
+    def on_pretrain_routine_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of each pretraining routine
+        """
+        self.run_callbacks('on_pretrain_routine_end', *args, **kwargs)
+
+    def on_train_start(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the start of each training
+        """
+        self.run_callbacks('on_train_start', *args, **kwargs)
+
+    def on_train_epoch_start(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the start of each training epoch
+        """
+        self.run_callbacks('on_train_epoch_start', *args, **kwargs)
+
+    def on_train_batch_start(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the start of each training batch
+        """
+        self.run_callbacks('on_train_batch_start', *args, **kwargs)
+
+    def optimizer_step(self, *args, **kwargs):
+        """
+        Fires all registered callbacks on each optimizer step
+        """
+        self.run_callbacks('optimizer_step', *args, **kwargs)
+
+    def on_before_zero_grad(self, *args, **kwargs):
+        """
+        Fires all registered callbacks before zero grad
+        """
+        self.run_callbacks('on_before_zero_grad', *args, **kwargs)
+
+    def on_train_batch_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of each training batch
+        """
+        self.run_callbacks('on_train_batch_end', *args, **kwargs)
+
+    def on_train_epoch_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of each training epoch
+        """
+        self.run_callbacks('on_train_epoch_end', *args, **kwargs)
+
+    def on_val_start(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the start of the validation
+        """
+        self.run_callbacks('on_val_start', *args, **kwargs)
+
+    def on_val_batch_start(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the start of each validation batch
+        """
+        self.run_callbacks('on_val_batch_start', *args, **kwargs)
+
+    def on_val_image_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of each val image
+        """
+        self.run_callbacks('on_val_image_end', *args, **kwargs)
+
+    def on_val_batch_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of each validation batch
+        """
+        self.run_callbacks('on_val_batch_end', *args, **kwargs)
+
+    def on_val_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of the validation
+        """
+        self.run_callbacks('on_val_end', *args, **kwargs)
+
+    def on_fit_epoch_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of each fit (train+val) epoch
+        """
+        self.run_callbacks('on_fit_epoch_end', *args, **kwargs)
+
+    def on_model_save(self, *args, **kwargs):
+        """
+        Fires all registered callbacks after each model save
+        """
+        self.run_callbacks('on_model_save', *args, **kwargs)
+
+    def on_train_end(self, *args, **kwargs):
+        """
+        Fires all registered callbacks at the end of training
+        """
+        self.run_callbacks('on_train_end', *args, **kwargs)
+
+    def teardown(self, *args, **kwargs):
+        """
+        Fires all registered callbacks before teardown
+        """
+        self.run_callbacks('teardown', *args, **kwargs)

utils/datasets.py

@@ -0,0 +1,1056 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Dataloaders and dataset utils
+"""
+
+import glob
+import hashlib
+import json
+import logging
+import os
+import random
+import shutil
+import time
+from itertools import repeat
+from multiprocessing.pool import ThreadPool, Pool
+from pathlib import Path
+from threading import Thread
+
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as F
+import yaml
+from PIL import Image, ExifTags
+from torch.utils.data import Dataset
+from tqdm import tqdm
+
+from utils.augmentations import Albumentations, augment_hsv, copy_paste, letterbox, mixup, random_perspective
+from utils.general import check_requirements, check_file, check_dataset, xywh2xyxy, xywhn2xyxy, xyxy2xywhn, \
+    xyn2xy, segments2boxes, clean_str
+from utils.torch_utils import torch_distributed_zero_first
+
+# Parameters
+HELP_URL = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
+IMG_FORMATS = ['bmp', 'jpg', 'jpeg', 'png', 'tif', 'tiff', 'dng', 'webp', 'mpo']  # acceptable image suffixes
+VID_FORMATS = ['mov', 'avi', 'mp4', 'mpg', 'mpeg', 'm4v', 'wmv', 'mkv']  # acceptable video suffixes
+NUM_THREADS = os.cpu_count()  # number of multiprocessing threads
+
+
+# Get orientation exif tag
+for orientation in ExifTags.TAGS.keys():
+    if ExifTags.TAGS[orientation] == 'Orientation':
+        break
+
+
+def get_hash(paths):
+    # Returns a single hash value of a list of paths (files or dirs)
+    size = sum(os.path.getsize(p) for p in paths if os.path.exists(p))  # sizes
+    h = hashlib.md5(str(size).encode())  # hash sizes
+    h.update(''.join(paths).encode())  # hash paths
+    return h.hexdigest()  # return hash
+
+
+def exif_size(img):
+    # Returns exif-corrected PIL size
+    s = img.size  # (width, height)
+    try:
+        rotation = dict(img._getexif().items())[orientation]
+        if rotation == 6:  # rotation 270
+            s = (s[1], s[0])
+        elif rotation == 8:  # rotation 90
+            s = (s[1], s[0])
+    except:
+        pass
+
+    return s
+
+
+def exif_transpose(image):
+    """
+    Transpose a PIL image accordingly if it has an EXIF Orientation tag.
+    From https://github.com/python-pillow/Pillow/blob/master/src/PIL/ImageOps.py
+
+    :param image: The image to transpose.
+    :return: An image.
+    """
+    exif = image.getexif()
+    orientation = exif.get(0x0112, 1)  # default 1
+    if orientation > 1:
+        method = {2: Image.FLIP_LEFT_RIGHT,
+                  3: Image.ROTATE_180,
+                  4: Image.FLIP_TOP_BOTTOM,
+                  5: Image.TRANSPOSE,
+                  6: Image.ROTATE_270,
+                  7: Image.TRANSVERSE,
+                  8: Image.ROTATE_90,
+                  }.get(orientation)
+        if method is not None:
+            image = image.transpose(method)
+            del exif[0x0112]
+            image.info["exif"] = exif.tobytes()
+    return image
+
+
+def create_dataloader(path, labels_dir, imgsz, batch_size, stride, single_cls=False, hyp=None, augment=False, cache=False, pad=0.0,
+                      rect=False, rank=-1, workers=8, image_weights=False, quad=False, prefix='',
+                      kp_flip=None, kp_bbox=None):
+    # Make sure only the first process in DDP process the dataset first, and the following others can use the cache
+    with torch_distributed_zero_first(rank):
+        dataset = LoadImagesAndLabels(path, labels_dir, imgsz, batch_size,
+                                      augment=augment,  # augment images
+                                      hyp=hyp,  # augmentation hyperparameters
+                                      rect=rect,  # rectangular training
+                                      cache_images=cache,
+                                      single_cls=single_cls,
+                                      stride=int(stride),
+                                      pad=pad,
+                                      image_weights=image_weights,
+                                      prefix=prefix,
+                                      kp_flip=kp_flip,
+                                      kp_bbox=kp_bbox)
+
+        # for i in range(10):
+        #     dataset.__getitem__(i)
+        # import sys
+        # sys.exit()
+
+    batch_size = min(batch_size, len(dataset))
+    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, workers])  # number of workers
+    sampler = torch.utils.data.distributed.DistributedSampler(dataset) if rank != -1 else None
+    loader = torch.utils.data.DataLoader if image_weights else InfiniteDataLoader
+    # Use torch.utils.data.DataLoader() if dataset.properties will update during training else InfiniteDataLoader()
+    dataloader = loader(dataset,
+                        batch_size=batch_size,
+                        num_workers=nw,
+                        sampler=sampler,
+                        pin_memory=True,
+                        collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn)
+    return dataloader, dataset
+
+
+class InfiniteDataLoader(torch.utils.data.dataloader.DataLoader):
+    """ Dataloader that reuses workers
+
+    Uses same syntax as vanilla DataLoader
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler))
+        self.iterator = super().__iter__()
+
+    def __len__(self):
+        return len(self.batch_sampler.sampler)
+
+    def __iter__(self):
+        for i in range(len(self)):
+            yield next(self.iterator)
+
+
+class _RepeatSampler(object):
+    """ Sampler that repeats forever
+
+    Args:
+        sampler (Sampler)
+    """
+
+    def __init__(self, sampler):
+        self.sampler = sampler
+
+    def __iter__(self):
+        while True:
+            yield from iter(self.sampler)
+
+
+class LoadImages:  # for inference
+    def __init__(self, path, img_size=640, stride=32, auto=True):
+        p = str(Path(path).absolute())  # os-agnostic absolute path
+        if p.endswith('.txt'):
+            with open(p, 'r') as f:
+                files = f.readlines()
+            files = [l.strip() for l in files]
+        elif '*' in p:
+            files = sorted(glob.glob(p, recursive=True))  # glob
+        elif os.path.isdir(p):
+            files = sorted(glob.glob(os.path.join(p, '*.*')))  # dir
+        elif os.path.isfile(p):
+            files = [p]  # files
+        else:
+            raise Exception(f'ERROR: {p} does not exist')
+
+        images = [x for x in files if x.split('.')[-1].lower() in IMG_FORMATS]
+        videos = [x for x in files if x.split('.')[-1].lower() in VID_FORMATS]
+        ni, nv = len(images), len(videos)
+
+        self.img_size = img_size
+        self.stride = stride
+        self.files = images + videos
+        self.nf = ni + nv  # number of files
+        self.video_flag = [False] * ni + [True] * nv
+        self.mode = 'image'
+        self.auto = auto
+        if any(videos):
+            self.new_video(videos[0])  # new video
+        else:
+            self.cap = None
+        assert self.nf > 0, f'No images or videos found in {p}. ' \
+                            f'Supported formats are:\nimages: {IMG_FORMATS}\nvideos: {VID_FORMATS}'
+
+    def __iter__(self):
+        self.count = 0
+        return self
+
+    def __next__(self):
+        if self.count == self.nf:
+            raise StopIteration
+        path = self.files[self.count]
+
+        if self.video_flag[self.count]:
+            # Read video
+            self.mode = 'video'
+            ret_val, img0 = self.cap.read()
+            if not ret_val:
+                self.count += 1
+                self.cap.release()
+                if self.count == self.nf:  # last video
+                    raise StopIteration
+                else:
+                    path = self.files[self.count]
+                    self.new_video(path)
+                    ret_val, img0 = self.cap.read()
+
+            self.frame += 1
+            # print(f'video {self.count + 1}/{self.nf} ({self.frame}/{self.frames}) {path}: ', end='')
+
+        else:
+            # Read image
+            self.count += 1
+            img0 = cv2.imread(path)  # BGR
+            assert img0 is not None, 'Image Not Found ' + path
+            print(f'image {self.count}/{self.nf} {path}: ', end='')
+
+        # Padded resize
+        img = letterbox(img0, self.img_size, stride=self.stride, auto=self.auto)[0]
+
+        # Convert
+        img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
+        img = np.ascontiguousarray(img)
+
+        return path, img, img0, self.cap
+
+    def new_video(self, path):
+        self.frame = 0
+        self.cap = cv2.VideoCapture(path)
+        self.frames = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+    def __len__(self):
+        return self.nf  # number of files
+
+
+class LoadWebcam:  # for inference
+    def __init__(self, pipe='0', img_size=640, stride=32):
+        self.img_size = img_size
+        self.stride = stride
+        self.pipe = eval(pipe) if pipe.isnumeric() else pipe
+        self.cap = cv2.VideoCapture(self.pipe)  # video capture object
+        self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3)  # set buffer size
+
+    def __iter__(self):
+        self.count = -1
+        return self
+
+    def __next__(self):
+        self.count += 1
+        if cv2.waitKey(1) == ord('q'):  # q to quit
+            self.cap.release()
+            cv2.destroyAllWindows()
+            raise StopIteration
+
+        # Read frame
+        ret_val, img0 = self.cap.read()
+        img0 = cv2.flip(img0, 1)  # flip left-right
+
+        # Print
+        assert ret_val, f'Camera Error {self.pipe}'
+        img_path = 'webcam.jpg'
+        print(f'webcam {self.count}: ', end='')
+
+        # Padded resize
+        img = letterbox(img0, self.img_size, stride=self.stride)[0]
+
+        # Convert
+        img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
+        img = np.ascontiguousarray(img)
+
+        return img_path, img, img0, None
+
+    def __len__(self):
+        return 0
+
+
+class LoadStreams:  # multiple IP or RTSP cameras
+    def __init__(self, sources='streams.txt', img_size=640, stride=32, auto=True):
+        self.mode = 'stream'
+        self.img_size = img_size
+        self.stride = stride
+
+        if os.path.isfile(sources):
+            with open(sources, 'r') as f:
+                sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())]
+        else:
+            sources = [sources]
+
+        n = len(sources)
+        self.imgs, self.fps, self.frames, self.threads = [None] * n, [0] * n, [0] * n, [None] * n
+        self.sources = [clean_str(x) for x in sources]  # clean source names for later
+        self.auto = auto
+        for i, s in enumerate(sources):  # index, source
+            # Start thread to read frames from video stream
+            print(f'{i + 1}/{n}: {s}... ', end='')
+            if 'youtube.com/' in s or 'youtu.be/' in s:  # if source is YouTube video
+                check_requirements(('pafy', 'youtube_dl'))
+                import pafy
+                s = pafy.new(s).getbest(preftype="mp4").url  # YouTube URL
+            s = eval(s) if s.isnumeric() else s  # i.e. s = '0' local webcam
+            cap = cv2.VideoCapture(s)
+            assert cap.isOpened(), f'Failed to open {s}'
+            w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+            h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+            self.fps[i] = max(cap.get(cv2.CAP_PROP_FPS) % 100, 0) or 30.0  # 30 FPS fallback
+            self.frames[i] = max(int(cap.get(cv2.CAP_PROP_FRAME_COUNT)), 0) or float('inf')  # infinite stream fallback
+
+            _, self.imgs[i] = cap.read()  # guarantee first frame
+            self.threads[i] = Thread(target=self.update, args=([i, cap]), daemon=True)
+            print(f" success ({self.frames[i]} frames {w}x{h} at {self.fps[i]:.2f} FPS)")
+            self.threads[i].start()
+        print('')  # newline
+
+        # check for common shapes
+        s = np.stack([letterbox(x, self.img_size, stride=self.stride, auto=self.auto)[0].shape for x in self.imgs])
+        self.rect = np.unique(s, axis=0).shape[0] == 1  # rect inference if all shapes equal
+        if not self.rect:
+            print('WARNING: Different stream shapes detected. For optimal performance supply similarly-shaped streams.')
+
+    def update(self, i, cap):
+        # Read stream `i` frames in daemon thread
+        n, f, read = 0, self.frames[i], 1  # frame number, frame array, inference every 'read' frame
+        while cap.isOpened() and n < f:
+            n += 1
+            # _, self.imgs[index] = cap.read()
+            cap.grab()
+            if n % read == 0:
+                success, im = cap.retrieve()
+                self.imgs[i] = im if success else self.imgs[i] * 0
+            time.sleep(1 / self.fps[i])  # wait time
+
+    def __iter__(self):
+        self.count = -1
+        return self
+
+    def __next__(self):
+        self.count += 1
+        if not all(x.is_alive() for x in self.threads) or cv2.waitKey(1) == ord('q'):  # q to quit
+            cv2.destroyAllWindows()
+            raise StopIteration
+
+        # Letterbox
+        img0 = self.imgs.copy()
+        img = [letterbox(x, self.img_size, stride=self.stride, auto=self.rect and self.auto)[0] for x in img0]
+
+        # Stack
+        img = np.stack(img, 0)
+
+        # Convert
+        img = img[..., ::-1].transpose((0, 3, 1, 2))  # BGR to RGB, BHWC to BCHW
+        img = np.ascontiguousarray(img)
+
+        return self.sources, img, img0, None
+
+    def __len__(self):
+        return len(self.sources)  # 1E12 frames = 32 streams at 30 FPS for 30 years
+
+
+def img2label_paths(img_paths, image_dir='images', labels_dir='labels'):
+    return [os.path.splitext(s.replace(image_dir, labels_dir))[0] + '.txt' for s in img_paths]
+
+
+class LoadImagesAndLabels(Dataset):  # for training/testing
+    def __init__(self, path, labels_dir='labels', img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
+                 cache_images=False, single_cls=False, stride=32, pad=0.0, prefix='',
+                 kp_flip=None, kp_bbox=None):
+        self.labels_dir = labels_dir
+        self.img_size = img_size
+        self.augment = augment
+        self.hyp = hyp
+        self.image_weights = image_weights
+        self.rect = False if image_weights else rect
+        self.mosaic = self.augment and not self.rect  # load 4 images at a time into a mosaic (only during training)
+        self.mosaic_border = [-img_size // 2, -img_size // 2]
+        self.stride = stride
+        self.path = path
+        self.albumentations = Albumentations() if augment else None
+        self.kp_flip = kp_flip
+        self.kp_bbox = kp_bbox
+        self.num_coords = len(kp_flip) * 2
+
+        if self.kp_flip:
+            self.obj_flip = {0: 0}
+            for i in range(len(self.kp_flip)):
+                self.obj_flip[i + 1] = self.kp_flip[i] + 1
+        else:
+            self.obj_flip = None
+
+        try:
+            f = []  # image files
+            for p in path if isinstance(path, list) else [path]:
+                p = Path(p)  # os-agnostic
+                if p.is_dir():  # dir
+                    f += glob.glob(str(p / '**' / '*.*'), recursive=True)
+                    # f = list(p.rglob('**/*.*'))  # pathlib
+                elif p.is_file():  # file
+                    with open(p, 'r') as t:
+                        t = t.read().strip().splitlines()
+                        parent = str(p.parent) + os.sep
+                        f += [x.replace('./', parent) if x.startswith('./') else x for x in t]  # local to global path
+                        # f += [p.parent / x.lstrip(os.sep) for x in t]  # local to global path (pathlib)
+                else:
+                    raise Exception(f'{prefix}{p} does not exist')
+            self.img_files = sorted([x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in IMG_FORMATS])
+            # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in img_formats])  # pathlib
+            assert self.img_files, f'{prefix}No images found'
+        except Exception as e:
+            raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {HELP_URL}')
+
+        # Check cache
+        self.label_files = img2label_paths(self.img_files, labels_dir=self.labels_dir)  # labels
+        cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache')
+        try:
+            cache, exists = np.load(cache_path, allow_pickle=True).item(), True  # load dict
+            assert cache['version'] == 0.4 and cache['hash'] == get_hash(self.label_files + self.img_files)
+        except:
+            cache, exists = self.cache_labels(cache_path, prefix), False  # cache
+
+        # Display cache
+        nf, nm, ne, nc, n = cache.pop('results')  # found, missing, empty, corrupted, total
+        if exists:
+            d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupted"
+            tqdm(None, desc=prefix + d, total=n, initial=n)  # display cache results
+            if cache['msgs']:
+                logging.info('\n'.join(cache['msgs']))  # display warnings
+        assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {HELP_URL}'
+
+        # Read cache
+        [cache.pop(k) for k in ('hash', 'version', 'msgs')]  # remove items
+        labels, shapes, self.segments = zip(*cache.values())
+        self.labels = list(labels)
+        self.shapes = np.array(shapes, dtype=np.float64)
+        self.img_files = list(cache.keys())  # update
+        self.label_files = img2label_paths(cache.keys(), labels_dir=self.labels_dir)  # update
+        if single_cls:
+            for x in self.labels:
+                x[:, 0] = 0
+
+        n = len(shapes)  # number of images
+        bi = np.floor(np.arange(n) / batch_size).astype(np.int)  # batch index
+        nb = bi[-1] + 1  # number of batches
+        self.batch = bi  # batch index of image
+        self.n = n
+        self.indices = range(n)
+
+        # Rectangular Training
+        if self.rect:
+            # Sort by aspect ratio
+            s = self.shapes  # wh
+            ar = s[:, 1] / s[:, 0]  # aspect ratio
+            irect = ar.argsort()
+            self.img_files = [self.img_files[i] for i in irect]
+            self.label_files = [self.label_files[i] for i in irect]
+            self.labels = [self.labels[i] for i in irect]
+            self.shapes = s[irect]  # wh
+            ar = ar[irect]
+
+            # Set training image shapes
+            shapes = [[1, 1]] * nb
+            for i in range(nb):
+                ari = ar[bi == i]
+                mini, maxi = ari.min(), ari.max()
+                if maxi < 1:
+                    shapes[i] = [maxi, 1]
+                elif mini > 1:
+                    shapes[i] = [1, 1 / mini]
+
+            self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride
+
+        # Cache images into memory for faster training (WARNING: large datasets may exceed system RAM)
+        self.imgs, self.img_npy = [None] * n, [None] * n
+        if cache_images:
+            if cache_images == 'disk':
+                self.im_cache_dir = Path(Path(self.img_files[0]).parent.as_posix() + '_npy')
+                self.img_npy = [self.im_cache_dir / Path(f).with_suffix('.npy').name for f in self.img_files]
+                self.im_cache_dir.mkdir(parents=True, exist_ok=True)
+            gb = 0  # Gigabytes of cached images
+            self.img_hw0, self.img_hw = [None] * n, [None] * n
+            results = ThreadPool(NUM_THREADS).imap(lambda x: load_image(*x), zip(repeat(self), range(n)))
+            pbar = tqdm(enumerate(results), total=n)
+            for i, x in pbar:
+                if cache_images == 'disk':
+                    if not self.img_npy[i].exists():
+                        np.save(self.img_npy[i].as_posix(), x[0])
+                    gb += self.img_npy[i].stat().st_size
+                else:
+                    self.imgs[i], self.img_hw0[i], self.img_hw[i] = x  # im, hw_orig, hw_resized = load_image(self, i)
+                    gb += self.imgs[i].nbytes
+                pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB {cache_images})'
+            pbar.close()
+
+    def cache_labels(self, path=Path('./labels.cache'), prefix=''):
+        # Cache dataset labels, check images and read shapes
+        x = {}  # dict
+        nm, nf, ne, nc, msgs = 0, 0, 0, 0, []  # number missing, found, empty, corrupt, messages
+        desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels..."
+        with Pool(NUM_THREADS) as pool:
+            pbar = tqdm(pool.imap_unordered(verify_image_label, zip(self.img_files, self.label_files, repeat(prefix), repeat(self.num_coords))),
+                        desc=desc, total=len(self.img_files))
+            for im_file, l, shape, segments, nm_f, nf_f, ne_f, nc_f, msg in pbar:
+                nm += nm_f
+                nf += nf_f
+                ne += ne_f
+                nc += nc_f
+                if im_file:
+                    x[im_file] = [l, shape, segments]
+                if msg:
+                    msgs.append(msg)
+                pbar.desc = f"{desc}{nf} found, {nm} missing, {ne} empty, {nc} corrupted"
+
+        pbar.close()
+        if msgs:
+            logging.info('\n'.join(msgs))
+        if nf == 0:
+            logging.info(f'{prefix}WARNING: No labels found in {path}. See {HELP_URL}')
+        x['hash'] = get_hash(self.label_files + self.img_files)
+        x['results'] = nf, nm, ne, nc, len(self.img_files)
+        x['msgs'] = msgs  # warnings
+        x['version'] = 0.4  # cache version
+        try:
+            np.save(path, x)  # save cache for next time
+            path.with_suffix('.cache.npy').rename(path)  # remove .npy suffix
+            logging.info(f'{prefix}New cache created: {path}')
+        except Exception as e:
+            logging.info(f'{prefix}WARNING: Cache directory {path.parent} is not writeable: {e}')  # path not writeable
+        return x
+
+    def __len__(self):
+        return len(self.img_files)
+
+    # def __iter__(self):
+    #     self.count = -1
+    #     print('ran dataset iter')
+    #     #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF)
+    #     return self
+
+    def __getitem__(self, index):
+        index = self.indices[index]  # linear, shuffled, or image_weights
+
+        hyp = self.hyp
+        mosaic = self.mosaic and random.random() < hyp['mosaic']
+        if mosaic:
+            # Load mosaic
+            img, labels = load_mosaic(self, index, kp_bbox=self.kp_bbox)
+            shapes = None
+
+            # MixUp augmentation
+            if random.random() < hyp['mixup']:
+                img, labels = mixup(img, labels, *load_mosaic(self, random.randint(0, self.n - 1)))
+
+        else:
+            # Load image
+            img, (h0, w0), (h, w) = load_image(self, index)
+
+            # Letterbox
+            shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size  # final letterboxed shape
+            img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
+            shapes = (h0, w0), ((h / h0, w / w0), pad)  # for COCO mAP rescaling
+
+            labels = self.labels[index].copy()
+            if labels.size:  # normalized xywh to pixel xyxy format
+                labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])
+
+            if self.augment:
+                img, labels = random_perspective(img, labels,
+                                                 degrees=hyp['degrees'],
+                                                 translate=hyp['translate'],
+                                                 scale=hyp['scale'],
+                                                 shear=hyp['shear'],
+                                                 perspective=hyp['perspective'],
+                                                 kp_bbox=self.kp_bbox)
+
+        nl = len(labels)  # number of labels
+        if nl:
+            labels[:, 1:] = xyxy2xywhn(labels[:, 1:], w=img.shape[1], h=img.shape[0], clip=True, eps=1E-3)
+
+        if self.augment:
+            # Albumentations
+            img, labels = self.albumentations(img, labels)
+            nl = len(labels)  # update after albumentations
+
+            # HSV color-space
+            augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])
+
+            # Flip up-down
+            if random.random() < hyp['flipud']:
+                img = np.flipud(img)
+                if nl:
+                    labels[:, 2] = 1 - labels[:, 2]
+
+            # Flip left-right
+            if random.random() < hyp['fliplr']:
+                img = np.fliplr(img)
+                if nl:
+                    labels[:, 1] = 1 - labels[:, 1]
+
+                    if self.kp_flip and labels.shape[1] > 5:
+                        labels[:, 5::3] = 1 - labels[:, 5::3]  # flip keypoints in person object
+                        keypoints = labels[:, 5:].reshape(nl, -1, 3)
+                        keypoints = keypoints[:, self.kp_flip]  # reorder left / right keypoints
+                        labels[:, 5:] = keypoints.reshape(nl, -1)
+
+                    if self.obj_flip:
+                        for i, cls in enumerate(labels[:, 0]):
+                            labels[i, 0] = self.obj_flip[labels[i, 0]]
+
+            # Cutouts
+            # labels = cutout(img, labels, p=0.5)
+
+        # img_h, img_w = img.shape[:2]
+        # img = img.copy()
+        # person_obj = labels[labels[:, 0] == 0]
+        # for lbl in person_obj:
+        #     xc, yc, w, h = lbl[1:5].copy()
+        #     pt1 = (int((xc - w / 2) * img_w), int((yc - h / 2) * img_h))
+        #     pt2 = (int((xc + w / 2) * img_w), int((yc + h / 2) * img_h))
+        #     cv2.rectangle(img, pt1, pt2, (255, 0, 255), thickness=2)
+        #
+        #     kp = lbl[5:]
+        #     kp = np.array(kp).reshape(-1, 3)
+        #     kp[:, 0] = kp[:, 0] * img_w
+        #     kp[:, 1] = kp[:, 1] * img_h
+        #     for i, (x, y, v) in enumerate(kp):
+        #         if v:
+        #             if i in COCO_KP_LEFT:
+        #                 color = (0, 255, 255)
+        #             else:
+        #                 color = (255, 255, 0)
+        #             cv2.circle(img, (int(round(x)), int(round(y))), 2, color, thickness=2)
+        #             # cv2.putText(img, COCO_KP_NAMES_SHORT[i], (int(round(x + 10)), int(round(y + 10))),
+        #             #             cv2.FONT_HERSHEY_PLAIN, 1, (255, 255, 0), thickness=1)
+        # cv2.imshow('', img)
+        # cv2.waitKey(0)
+        # cv2.destroyAllWindows()
+
+        labels_out = torch.zeros((nl, labels.shape[-1] + 1))
+        if nl:
+            labels_out[:, 1:] = torch.from_numpy(labels)
+        # Convert
+        img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
+        img = np.ascontiguousarray(img)
+
+        return torch.from_numpy(img), labels_out, self.img_files[index], shapes
+
+    @staticmethod
+    def collate_fn(batch):
+        img, label, path, shapes = zip(*batch)  # transposed
+        for i, l in enumerate(label):
+            l[:, 0] = i  # add target image index for build_targets()
+        return torch.stack(img, 0), torch.cat(label, 0), path, shapes
+
+    @staticmethod
+    def collate_fn4(batch):
+        img, label, path, shapes = zip(*batch)  # transposed
+        n = len(shapes) // 4
+        img4, label4, path4, shapes4 = [], [], path[:n], shapes[:n]
+
+        ho = torch.tensor([[0., 0, 0, 1, 0, 0]])
+        wo = torch.tensor([[0., 0, 1, 0, 0, 0]])
+        s = torch.tensor([[1, 1, .5, .5, .5, .5]])  # scale
+        for i in range(n):  # zidane torch.zeros(16,3,720,1280)  # BCHW
+            i *= 4
+            if random.random() < 0.5:
+                im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2., mode='bilinear', align_corners=False)[
+                    0].type(img[i].type())
+                l = label[i]
+            else:
+                im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2)
+                l = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s
+            img4.append(im)
+            label4.append(l)
+
+        for i, l in enumerate(label4):
+            l[:, 0] = i  # add target image index for build_targets()
+
+        return torch.stack(img4, 0), torch.cat(label4, 0), path4, shapes4
+
+
+# Ancillary functions --------------------------------------------------------------------------------------------------
+def load_image(self, i):
+    # loads 1 image from dataset index 'i', returns im, original hw, resized hw
+    im = self.imgs[i]
+    if im is None:  # not cached in ram
+        npy = self.img_npy[i]
+        if npy and npy.exists():  # load npy
+            im = np.load(npy)
+        else:  # read image
+            path = self.img_files[i]
+            im = cv2.imread(path)  # BGR
+            assert im is not None, 'Image Not Found ' + path
+        h0, w0 = im.shape[:2]  # orig hw
+        r = self.img_size / max(h0, w0)  # ratio
+        if r != 1:  # if sizes are not equal
+            im = cv2.resize(im, (int(w0 * r), int(h0 * r)),
+                            interpolation=cv2.INTER_AREA if r < 1 and not self.augment else cv2.INTER_LINEAR)
+        return im, (h0, w0), im.shape[:2]  # im, hw_original, hw_resized
+    else:
+        return self.imgs[i], self.img_hw0[i], self.img_hw[i]  # im, hw_original, hw_resized
+
+
+def load_mosaic(self, index, kp_bbox=None):
+    # loads images in a 4-mosaic
+
+    labels4, segments4 = [], []
+    s = self.img_size
+    yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border]  # mosaic center x, y
+    indices = [index] + random.choices(self.indices, k=3)  # 3 additional image indices
+    for i, index in enumerate(indices):
+        # Load image
+        img, _, (h, w) = load_image(self, index)
+
+        # place img in img4
+        if i == 0:  # top left
+            img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
+            x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
+            x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
+        elif i == 1:  # top right
+            x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
+            x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
+        elif i == 2:  # bottom left
+            x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
+            x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
+        elif i == 3:  # bottom right
+            x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
+            x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
+
+        img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
+        padw = x1a - x1b
+        padh = y1a - y1b
+
+        # Labels
+        labels, segments = self.labels[index].copy(), self.segments[index].copy()
+        if labels.size:
+            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh)  # normalized xywh to pixel xyxy format
+            segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
+        labels4.append(labels)
+        segments4.extend(segments)
+
+    # Concat/clip labels
+    labels4 = np.concatenate(labels4, 0)
+    for x in (labels4[:, 1:], *segments4):
+        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
+    # img4, labels4 = replicate(img4, labels4)  # replicate
+
+    # Augment
+    img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste'])
+    img4, labels4 = random_perspective(img4, labels4, segments4,
+                                       degrees=self.hyp['degrees'],
+                                       translate=self.hyp['translate'],
+                                       scale=self.hyp['scale'],
+                                       shear=self.hyp['shear'],
+                                       perspective=self.hyp['perspective'],
+                                       border=self.mosaic_border,
+                                       kp_bbox=kp_bbox)  # border to remove
+
+    return img4, labels4
+
+
+def load_mosaic9(self, index):
+    # loads images in a 9-mosaic
+
+    labels9, segments9 = [], []
+    s = self.img_size
+    indices = [index] + random.choices(self.indices, k=8)  # 8 additional image indices
+    for i, index in enumerate(indices):
+        # Load image
+        img, _, (h, w) = load_image(self, index)
+
+        # place img in img9
+        if i == 0:  # center
+            img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
+            h0, w0 = h, w
+            c = s, s, s + w, s + h  # xmin, ymin, xmax, ymax (base) coordinates
+        elif i == 1:  # top
+            c = s, s - h, s + w, s
+        elif i == 2:  # top right
+            c = s + wp, s - h, s + wp + w, s
+        elif i == 3:  # right
+            c = s + w0, s, s + w0 + w, s + h
+        elif i == 4:  # bottom right
+            c = s + w0, s + hp, s + w0 + w, s + hp + h
+        elif i == 5:  # bottom
+            c = s + w0 - w, s + h0, s + w0, s + h0 + h
+        elif i == 6:  # bottom left
+            c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
+        elif i == 7:  # left
+            c = s - w, s + h0 - h, s, s + h0
+        elif i == 8:  # top left
+            c = s - w, s + h0 - hp - h, s, s + h0 - hp
+
+        padx, pady = c[:2]
+        x1, y1, x2, y2 = [max(x, 0) for x in c]  # allocate coords
+
+        # Labels
+        labels, segments = self.labels[index].copy(), self.segments[index].copy()
+        if labels.size:
+            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady)  # normalized xywh to pixel xyxy format
+            segments = [xyn2xy(x, w, h, padx, pady) for x in segments]
+        labels9.append(labels)
+        segments9.extend(segments)
+
+        # Image
+        img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:]  # img9[ymin:ymax, xmin:xmax]
+        hp, wp = h, w  # height, width previous
+
+    # Offset
+    yc, xc = [int(random.uniform(0, s)) for _ in self.mosaic_border]  # mosaic center x, y
+    img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s]
+
+    # Concat/clip labels
+    labels9 = np.concatenate(labels9, 0)
+    labels9[:, [1, 3]] -= xc
+    labels9[:, [2, 4]] -= yc
+    c = np.array([xc, yc])  # centers
+    segments9 = [x - c for x in segments9]
+
+    for x in (labels9[:, 1:], *segments9):
+        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
+    # img9, labels9 = replicate(img9, labels9)  # replicate
+
+    # Augment
+    img9, labels9 = random_perspective(img9, labels9, segments9,
+                                       degrees=self.hyp['degrees'],
+                                       translate=self.hyp['translate'],
+                                       scale=self.hyp['scale'],
+                                       shear=self.hyp['shear'],
+                                       perspective=self.hyp['perspective'],
+                                       border=self.mosaic_border)  # border to remove
+
+    return img9, labels9
+
+
+def create_folder(path='./new'):
+    # Create folder
+    if os.path.exists(path):
+        shutil.rmtree(path)  # delete output folder
+    os.makedirs(path)  # make new output folder
+
+
+def flatten_recursive(path='../datasets/coco128'):
+    # Flatten a recursive directory by bringing all files to top level
+    new_path = Path(path + '_flat')
+    create_folder(new_path)
+    for file in tqdm(glob.glob(str(Path(path)) + '/**/*.*', recursive=True)):
+        shutil.copyfile(file, new_path / Path(file).name)
+
+
+def extract_boxes(path='../datasets/coco128', labels_dir='labels'):  # from utils.datasets import *; extract_boxes()
+    # Convert detection dataset into classification dataset, with one directory per class
+    path = Path(path)  # images dir
+    shutil.rmtree(path / 'classifier') if (path / 'classifier').is_dir() else None  # remove existing
+    files = list(path.rglob('*.*'))
+    n = len(files)  # number of files
+    for im_file in tqdm(files, total=n):
+        if im_file.suffix[1:] in IMG_FORMATS:
+            # image
+            im = cv2.imread(str(im_file))[..., ::-1]  # BGR to RGB
+            h, w = im.shape[:2]
+
+            # labels
+            lb_file = Path(img2label_paths([str(im_file)], labels_dir=labels_dir)[0])
+            if Path(lb_file).exists():
+                with open(lb_file, 'r') as f:
+                    lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32)  # labels
+
+                for j, x in enumerate(lb):
+                    c = int(x[0])  # class
+                    f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg'  # new filename
+                    if not f.parent.is_dir():
+                        f.parent.mkdir(parents=True)
+
+                    b = x[1:] * [w, h, w, h]  # box
+                    # b[2:] = b[2:].max()  # rectangle to square
+                    b[2:] = b[2:] * 1.2 + 3  # pad
+                    b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int)
+
+                    b[[0, 2]] = np.clip(b[[0, 2]], 0, w)  # clip boxes outside of image
+                    b[[1, 3]] = np.clip(b[[1, 3]], 0, h)
+                    assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}'
+
+
+def autosplit(path='../datasets/coco128/images', weights=(0.9, 0.1, 0.0), annotated_only=False, labels_dir='labels_dir'):
+    """ Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files
+    Usage: from utils.datasets import *; autosplit()
+    Arguments
+        path:            Path to images directory
+        weights:         Train, val, test weights (list, tuple)
+        annotated_only:  Only use images with an annotated txt file
+    """
+    path = Path(path)  # images dir
+    files = sum([list(path.rglob(f"*.{img_ext}")) for img_ext in IMG_FORMATS], [])  # image files only
+    n = len(files)  # number of files
+    random.seed(0)  # for reproducibility
+    indices = random.choices([0, 1, 2], weights=weights, k=n)  # assign each image to a split
+
+    txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt']  # 3 txt files
+    [(path.parent / x).unlink(missing_ok=True) for x in txt]  # remove existing
+
+    print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only)
+    for i, img in tqdm(zip(indices, files), total=n):
+        if not annotated_only or Path(img2label_paths([str(img)], labels_dir='labels_dir')[0]).exists():  # check label
+            with open(path.parent / txt[i], 'a') as f:
+                f.write('./' + img.relative_to(path.parent).as_posix() + '\n')  # add image to txt file
+
+
+def verify_image_label(args):
+    # Verify one image-label pair
+    im_file, lb_file, prefix, num_coords = args
+    nm, nf, ne, nc, msg, segments = 0, 0, 0, 0, '', []  # number (missing, found, empty, corrupt), message, segments
+    try:
+        # verify images
+        im = Image.open(im_file)
+        im.verify()  # PIL verify
+        shape = exif_size(im)  # image size
+        assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels'
+        assert im.format.lower() in IMG_FORMATS, f'invalid image format {im.format}'
+        if im.format.lower() in ('jpg', 'jpeg'):
+            with open(im_file, 'rb') as f:
+                f.seek(-2, 2)
+                if f.read() != b'\xff\xd9':  # corrupt JPEG
+                    Image.open(im_file).save(im_file, format='JPEG', subsampling=0, quality=100)  # re-save image
+                    msg = f'{prefix}WARNING: corrupt JPEG restored and saved {im_file}'
+
+        # verify labels
+        if os.path.isfile(lb_file):
+            nf = 1  # label found
+            with open(lb_file, 'r') as f:
+                l = [x.split() for x in f.read().strip().splitlines() if len(x)]
+                # if any([len(x) > 8 for x in l]):  # is segment
+                #     classes = np.array([x[0] for x in l], dtype=np.float32)
+                #     segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in l]  # (cls, xy1...)
+                #     l = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1)  # (cls, xywh)
+                l = np.array(l, dtype=np.float32)
+            if len(l):
+                # assert l.shape[1] == 5, 'labels require 5 columns each'
+                assert (l >= 0).all(), 'negative labels'
+                # assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels'
+                # assert np.unique(l, axis=0).shape[0] == l.shape[0], 'duplicate labels'
+            else:
+                ne = 1  # label empty
+                l = np.zeros((0, 5 + num_coords * 3 // 2), dtype=np.float32)
+        else:
+            nm = 1  # label missing
+            l = np.zeros((0, 5 + num_coords * 3 // 2), dtype=np.float32)
+        return im_file, l, shape, segments, nm, nf, ne, nc, msg
+    except Exception as e:
+        nc = 1
+        msg = f'{prefix}WARNING: Ignoring corrupted image and/or label {im_file}: {e}'
+        return [None, None, None, None, nm, nf, ne, nc, msg]
+
+
+def dataset_stats(path='coco128.yaml', autodownload=False, verbose=False, profile=False, hub=False, labels_dir='labels'):
+    """ Return dataset statistics dictionary with images and instances counts per split per class
+    To run in parent directory: export PYTHONPATH="$PWD/yolov5"
+    Usage1: from utils.datasets import *; dataset_stats('coco128.yaml', autodownload=True)
+    Usage2: from utils.datasets import *; dataset_stats('../datasets/coco128_with_yaml.zip')
+    Arguments
+        path:           Path to data.yaml or data.zip (with data.yaml inside data.zip)
+        autodownload:   Attempt to download dataset if not found locally
+        verbose:        Print stats dictionary
+    """
+
+    def round_labels(labels):
+        # Update labels to integer class and 6 decimal place floats
+        return [[int(c), *[round(x, 4) for x in points]] for c, *points in labels]
+
+    def unzip(path):
+        # Unzip data.zip TODO: CONSTRAINT: path/to/abc.zip MUST unzip to 'path/to/abc/'
+        if str(path).endswith('.zip'):  # path is data.zip
+            assert Path(path).is_file(), f'Error unzipping {path}, file not found'
+            assert os.system(f'unzip -q {path} -d {path.parent}') == 0, f'Error unzipping {path}'
+            dir = path.with_suffix('')  # dataset directory
+            return True, str(dir), next(dir.rglob('*.yaml'))  # zipped, data_dir, yaml_path
+        else:  # path is data.yaml
+            return False, None, path
+
+    def hub_ops(f, max_dim=1920):
+        # HUB ops for 1 image 'f'
+        im = Image.open(f)
+        r = max_dim / max(im.height, im.width)  # ratio
+        if r < 1.0:  # image too large
+            im = im.resize((int(im.width * r), int(im.height * r)))
+        im.save(im_dir / Path(f).name, quality=75)  # save
+
+    zipped, data_dir, yaml_path = unzip(Path(path))
+    with open(check_file(yaml_path), errors='ignore') as f:
+        data = yaml.safe_load(f)  # data dict
+        if zipped:
+            data['path'] = data_dir  # TODO: should this be dir.resolve()?
+    check_dataset(data, autodownload)  # download dataset if missing
+    hub_dir = Path(data['path'] + ('-hub' if hub else ''))
+    stats = {'nc': data['nc'], 'names': data['names']}  # statistics dictionary
+    for split in 'train', 'val', 'test':
+        if data.get(split) is None:
+            stats[split] = None  # i.e. no test set
+            continue
+        x = []
+        dataset = LoadImagesAndLabels(data[split], labels_dir=labels_dir)  # load dataset
+        for label in tqdm(dataset.labels, total=dataset.n, desc='Statistics'):
+            x.append(np.bincount(label[:, 0].astype(int), minlength=data['nc']))
+        x = np.array(x)  # shape(128x80)
+        stats[split] = {'instance_stats': {'total': int(x.sum()), 'per_class': x.sum(0).tolist()},
+                        'image_stats': {'total': dataset.n, 'unlabelled': int(np.all(x == 0, 1).sum()),
+                                        'per_class': (x > 0).sum(0).tolist()},
+                        'labels': [{str(Path(k).name): round_labels(v.tolist())} for k, v in
+                                   zip(dataset.img_files, dataset.labels)]}
+
+        if hub:
+            im_dir = hub_dir / 'images'
+            im_dir.mkdir(parents=True, exist_ok=True)
+            for _ in tqdm(ThreadPool(NUM_THREADS).imap(hub_ops, dataset.img_files), total=dataset.n, desc='HUB Ops'):
+                pass
+
+    # Profile
+    stats_path = hub_dir / 'stats.json'
+    if profile:
+        for _ in range(1):
+            file = stats_path.with_suffix('.npy')
+            t1 = time.time()
+            np.save(file, stats)
+            t2 = time.time()
+            x = np.load(file, allow_pickle=True)
+            print(f'stats.npy times: {time.time() - t2:.3f}s read, {t2 - t1:.3f}s write')
+
+            file = stats_path.with_suffix('.json')
+            t1 = time.time()
+            with open(file, 'w') as f:
+                json.dump(stats, f)  # save stats *.json
+            t2 = time.time()
+            with open(file, 'r') as f:
+                x = json.load(f)  # load hyps dict
+            print(f'stats.json times: {time.time() - t2:.3f}s read, {t2 - t1:.3f}s write')
+
+    # Save, print and return
+    if hub:
+        print(f'Saving {stats_path.resolve()}...')
+        with open(stats_path, 'w') as f:
+            json.dump(stats, f)  # save stats.json
+    if verbose:
+        print(json.dumps(stats, indent=2, sort_keys=False))
+    return stats

utils/downloads.py

@@ -0,0 +1,149 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Download utils
+"""
+
+import os
+import platform
+import subprocess
+import time
+import urllib
+from pathlib import Path
+
+import requests
+import torch
+
+
+def gsutil_getsize(url=''):
+    # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du
+    s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8')
+    return eval(s.split(' ')[0]) if len(s) else 0  # bytes
+
+
+def safe_download(file, url, url2=None, min_bytes=1E0, error_msg=''):
+    # Attempts to download file from url or url2, checks and removes incomplete downloads < min_bytes
+    file = Path(file)
+    assert_msg = f"Downloaded file '{file}' does not exist or size is < min_bytes={min_bytes}"
+    try:  # url1
+        print(f'Downloading {url} to {file}...')
+        torch.hub.download_url_to_file(url, str(file))
+        assert file.exists() and file.stat().st_size > min_bytes, assert_msg  # check
+    except Exception as e:  # url2
+        file.unlink(missing_ok=True)  # remove partial downloads
+        print(f'ERROR: {e}\nRe-attempting {url2 or url} to {file}...')
+        os.system(f"curl -L '{url2 or url}' -o '{file}' --retry 3 -C -")  # curl download, retry and resume on fail
+    finally:
+        if not file.exists() or file.stat().st_size < min_bytes:  # check
+            file.unlink(missing_ok=True)  # remove partial downloads
+            print(f"ERROR: {assert_msg}\n{error_msg}")
+        print('')
+
+
+def attempt_download(file, repo='ultralytics/yolov5'):  # from utils.downloads import *; attempt_download()
+    # Attempt file download if does not exist
+    file = Path(str(file).strip().replace("'", ''))
+
+    if not file.exists():
+        # URL specified
+        name = Path(urllib.parse.unquote(str(file))).name  # decode '%2F' to '/' etc.
+        if str(file).startswith(('http:/', 'https:/')):  # download
+            url = str(file).replace(':/', '://')  # Pathlib turns :// -> :/
+            name = name.split('?')[0]  # parse authentication https://url.com/file.txt?auth...
+            safe_download(file=name, url=url, min_bytes=1E5)
+            return name
+
+        # GitHub assets
+        file.parent.mkdir(parents=True, exist_ok=True)  # make parent dir (if required)
+        try:
+            response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest').json()  # github api
+            assets = [x['name'] for x in response['assets']]  # release assets, i.e. ['yolov5s.pt', 'yolov5m.pt', ...]
+            tag = response['tag_name']  # i.e. 'v1.0'
+        except:  # fallback plan
+            assets = ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt',
+                      'yolov5s6.pt', 'yolov5m6.pt', 'yolov5l6.pt', 'yolov5x6.pt']
+            try:
+                tag = subprocess.check_output('git tag', shell=True, stderr=subprocess.STDOUT).decode().split()[-1]
+            except:
+                tag = 'v5.0'  # current release
+        tag = 'v5.0'  # download v5.0 models
+        if name in assets:
+            safe_download(file,
+                          url=f'https://github.com/{repo}/releases/download/{tag}/{name}',
+                          # url2=f'https://storage.googleapis.com/{repo}/ckpt/{name}',  # backup url (optional)
+                          min_bytes=1E5,
+                          error_msg=f'{file} missing, try downloading from https://github.com/{repo}/releases/')
+
+    return str(file)
+
+
+def gdrive_download(id='16TiPfZj7htmTyhntwcZyEEAejOUxuT6m', file='tmp.zip'):
+    # Downloads a file from Google Drive. from yolov5.utils.downloads import *; gdrive_download()
+    t = time.time()
+    file = Path(file)
+    cookie = Path('cookie')  # gdrive cookie
+    print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='')
+    file.unlink(missing_ok=True)  # remove existing file
+    cookie.unlink(missing_ok=True)  # remove existing cookie
+
+    # Attempt file download
+    out = "NUL" if platform.system() == "Windows" else "/dev/null"
+    os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}')
+    if os.path.exists('cookie'):  # large file
+        s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}'
+    else:  # small file
+        s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"'
+    r = os.system(s)  # execute, capture return
+    cookie.unlink(missing_ok=True)  # remove existing cookie
+
+    # Error check
+    if r != 0:
+        file.unlink(missing_ok=True)  # remove partial
+        print('Download error ')  # raise Exception('Download error')
+        return r
+
+    # Unzip if archive
+    if file.suffix == '.zip':
+        print('unzipping... ', end='')
+        os.system(f'unzip -q {file}')  # unzip
+        file.unlink()  # remove zip to free space
+
+    print(f'Done ({time.time() - t:.1f}s)')
+    return r
+
+
+def get_token(cookie="./cookie"):
+    with open(cookie) as f:
+        for line in f:
+            if "download" in line:
+                return line.split()[-1]
+    return ""
+
+# Google utils: https://cloud.google.com/storage/docs/reference/libraries ----------------------------------------------
+#
+#
+# def upload_blob(bucket_name, source_file_name, destination_blob_name):
+#     # Uploads a file to a bucket
+#     # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python
+#
+#     storage_client = storage.Client()
+#     bucket = storage_client.get_bucket(bucket_name)
+#     blob = bucket.blob(destination_blob_name)
+#
+#     blob.upload_from_filename(source_file_name)
+#
+#     print('File {} uploaded to {}.'.format(
+#         source_file_name,
+#         destination_blob_name))
+#
+#
+# def download_blob(bucket_name, source_blob_name, destination_file_name):
+#     # Uploads a blob from a bucket
+#     storage_client = storage.Client()
+#     bucket = storage_client.get_bucket(bucket_name)
+#     blob = bucket.blob(source_blob_name)
+#
+#     blob.download_to_filename(destination_file_name)
+#
+#     print('Blob {} downloaded to {}.'.format(
+#         source_blob_name,
+#         destination_file_name))

utils/general.py

@@ -0,0 +1,853 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+General utils
+"""
+
+import contextlib
+import glob
+import logging
+import math
+import os
+import platform
+import random
+import re
+import signal
+import time
+import urllib
+from itertools import repeat
+from multiprocessing.pool import ThreadPool
+from pathlib import Path
+from subprocess import check_output
+
+import cv2
+import numpy as np
+import pandas as pd
+import pkg_resources as pkg
+import torch
+import torchvision
+import yaml
+
+from utils.downloads import gsutil_getsize
+from utils.metrics import box_iou, fitness
+from utils.torch_utils import init_torch_seeds
+from utils.labels import write_kp_labels
+
+# Settings
+torch.set_printoptions(linewidth=320, precision=5, profile='long')
+np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format})  # format short g, %precision=5
+pd.options.display.max_columns = 10
+cv2.setNumThreads(0)  # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader)
+os.environ['NUMEXPR_MAX_THREADS'] = str(min(os.cpu_count(), 8))  # NumExpr max threads
+
+
+class Profile(contextlib.ContextDecorator):
+    # Usage: @Profile() decorator or 'with Profile():' context manager
+    def __enter__(self):
+        self.start = time.time()
+
+    def __exit__(self, type, value, traceback):
+        print(f'Profile results: {time.time() - self.start:.5f}s')
+
+
+class Timeout(contextlib.ContextDecorator):
+    # Usage: @Timeout(seconds) decorator or 'with Timeout(seconds):' context manager
+    def __init__(self, seconds, *, timeout_msg='', suppress_timeout_errors=True):
+        self.seconds = int(seconds)
+        self.timeout_message = timeout_msg
+        self.suppress = bool(suppress_timeout_errors)
+
+    def _timeout_handler(self, signum, frame):
+        raise TimeoutError(self.timeout_message)
+
+    def __enter__(self):
+        signal.signal(signal.SIGALRM, self._timeout_handler)  # Set handler for SIGALRM
+        signal.alarm(self.seconds)  # start countdown for SIGALRM to be raised
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        signal.alarm(0)  # Cancel SIGALRM if it's scheduled
+        if self.suppress and exc_type is TimeoutError:  # Suppress TimeoutError
+            return True
+
+
+def try_except(func):
+    # try-except function. Usage: @try_except decorator
+    def handler(*args, **kwargs):
+        try:
+            func(*args, **kwargs)
+        except Exception as e:
+            print(e)
+
+    return handler
+
+
+def methods(instance):
+    # Get class/instance methods
+    return [f for f in dir(instance) if callable(getattr(instance, f)) and not f.startswith("__")]
+
+
+def set_logging(rank=-1, verbose=True):
+    logging.basicConfig(
+        format="%(message)s",
+        level=logging.INFO if (verbose and rank in [-1, 0]) else logging.WARN)
+
+
+def init_seeds(seed=0):
+    # Initialize random number generator (RNG) seeds
+    random.seed(seed)
+    np.random.seed(seed)
+    init_torch_seeds(seed)
+
+
+def get_latest_run(search_dir='.'):
+    # Return path to most recent 'last.pt' in /runs (i.e. to --resume from)
+    last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True)
+    return max(last_list, key=os.path.getctime) if last_list else ''
+
+
+def is_docker():
+    # Is environment a Docker container?
+    return Path('/workspace').exists()  # or Path('/.dockerenv').exists()
+
+
+def is_colab():
+    # Is environment a Google Colab instance?
+    try:
+        import google.colab
+        return True
+    except Exception as e:
+        return False
+
+
+def is_pip():
+    # Is file in a pip package?
+    return 'site-packages' in Path(__file__).absolute().parts
+
+
+def is_ascii(s=''):
+    # Is string composed of all ASCII (no UTF) characters?
+    s = str(s)  # convert list, tuple, None, etc. to str
+    return len(s.encode().decode('ascii', 'ignore')) == len(s)
+
+
+def emojis(str=''):
+    # Return platform-dependent emoji-safe version of string
+    return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str
+
+
+def file_size(file):
+    # Return file size in MB
+    return Path(file).stat().st_size / 1e6
+
+
+def check_online():
+    # Check internet connectivity
+    import socket
+    try:
+        socket.create_connection(("1.1.1.1", 443), 5)  # check host accessibility
+        return True
+    except OSError:
+        return False
+
+
+@try_except
+def check_git_status():
+    # Recommend 'git pull' if code is out of date
+    msg = ', for updates see https://github.com/ultralytics/yolov5'
+    print(colorstr('github: '), end='')
+    assert Path('.git').exists(), 'skipping check (not a git repository)' + msg
+    assert not is_docker(), 'skipping check (Docker image)' + msg
+    assert check_online(), 'skipping check (offline)' + msg
+
+    cmd = 'git fetch && git config --get remote.origin.url'
+    url = check_output(cmd, shell=True, timeout=5).decode().strip().rstrip('.git')  # git fetch
+    branch = check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip()  # checked out
+    n = int(check_output(f'git rev-list {branch}..origin/master --count', shell=True))  # commits behind
+    if n > 0:
+        s = f"⚠️ YOLOv5 is out of date by {n} commit{'s' * (n > 1)}. Use `git pull` or `git clone {url}` to update."
+    else:
+        s = f'up to date with {url} ✅'
+    print(emojis(s))  # emoji-safe
+
+
+def check_python(minimum='3.6.2'):
+    # Check current python version vs. required python version
+    check_version(platform.python_version(), minimum, name='Python ')
+
+
+def check_version(current='0.0.0', minimum='0.0.0', name='version ', pinned=False):
+    # Check version vs. required version
+    current, minimum = (pkg.parse_version(x) for x in (current, minimum))
+    result = (current == minimum) if pinned else (current >= minimum)
+    assert result, f'{name}{minimum} required by YOLOv5, but {name}{current} is currently installed'
+
+
+@try_except
+def check_requirements(requirements='requirements.txt', exclude=(), install=True):
+    # Check installed dependencies meet requirements (pass *.txt file or list of packages)
+    prefix = colorstr('red', 'bold', 'requirements:')
+    check_python()  # check python version
+    if isinstance(requirements, (str, Path)):  # requirements.txt file
+        file = Path(requirements)
+        assert file.exists(), f"{prefix} {file.resolve()} not found, check failed."
+        requirements = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements(file.open()) if x.name not in exclude]
+    else:  # list or tuple of packages
+        requirements = [x for x in requirements if x not in exclude]
+
+    n = 0  # number of packages updates
+    for r in requirements:
+        try:
+            pkg.require(r)
+        except Exception as e:  # DistributionNotFound or VersionConflict if requirements not met
+            s = f"{prefix} {r} not found and is required by YOLOv5"
+            if install:
+                print(f"{s}, attempting auto-update...")
+                try:
+                    assert check_online(), f"'pip install {r}' skipped (offline)"
+                    print(check_output(f"pip install '{r}'", shell=True).decode())
+                    n += 1
+                except Exception as e:
+                    print(f'{prefix} {e}')
+            else:
+                print(f'{s}. Please install and rerun your command.')
+
+    if n:  # if packages updated
+        source = file.resolve() if 'file' in locals() else requirements
+        s = f"{prefix} {n} package{'s' * (n > 1)} updated per {source}\n" \
+            f"{prefix} ⚠️ {colorstr('bold', 'Restart runtime or rerun command for updates to take effect')}\n"
+        print(emojis(s))
+
+
+def check_img_size(imgsz, s=32, floor=0):
+    # Verify image size is a multiple of stride s in each dimension
+    if isinstance(imgsz, int):  # integer i.e. img_size=640
+        new_size = max(make_divisible(imgsz, int(s)), floor)
+    else:  # list i.e. img_size=[640, 480]
+        new_size = [max(make_divisible(x, int(s)), floor) for x in imgsz]
+    if new_size != imgsz:
+        print(f'WARNING: --img-size {imgsz} must be multiple of max stride {s}, updating to {new_size}')
+    return new_size
+
+
+def check_imshow():
+    # Check if environment supports image displays
+    try:
+        assert not is_docker(), 'cv2.imshow() is disabled in Docker environments'
+        assert not is_colab(), 'cv2.imshow() is disabled in Google Colab environments'
+        cv2.imshow('test', np.zeros((1, 1, 3)))
+        cv2.waitKey(1)
+        cv2.destroyAllWindows()
+        cv2.waitKey(1)
+        return True
+    except Exception as e:
+        print(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}')
+        return False
+
+
+def check_file(file):
+    # Search/download file (if necessary) and return path
+    file = str(file)  # convert to str()
+    if Path(file).is_file() or file == '':  # exists
+        return file
+    elif file.startswith(('http:/', 'https:/')):  # download
+        url = str(Path(file)).replace(':/', '://')  # Pathlib turns :// -> :/
+        file = Path(urllib.parse.unquote(file)).name.split('?')[0]  # '%2F' to '/', split https://url.com/file.txt?auth
+        print(f'Downloading {url} to {file}...')
+        torch.hub.download_url_to_file(url, file)
+        assert Path(file).exists() and Path(file).stat().st_size > 0, f'File download failed: {url}'  # check
+        return file
+    else:  # search
+        files = glob.glob('./**/' + file, recursive=True)  # find file
+        assert len(files), f'File not found: {file}'  # assert file was found
+        assert len(files) == 1, f"Multiple files match '{file}', specify exact path: {files}"  # assert unique
+        return files[0]  # return file
+
+
+def check_dataset(data, autodownload=True):
+    # Download and/or unzip dataset if not found locally
+    # Usage: https://github.com/ultralytics/yolov5/releases/download/v1.0/coco128_with_yaml.zip
+
+    # Download (optional)
+    extract_dir = ''
+    if isinstance(data, (str, Path)) and str(data).endswith('.zip'):  # i.e. gs://bucket/dir/coco128.zip
+        download(data, dir='../datasets', unzip=True, delete=False, curl=False, threads=1)
+        data = next((Path('../datasets') / Path(data).stem).rglob('*.yaml'))
+        extract_dir, autodownload = data.parent, False
+
+    # Read yaml (optional)
+    if isinstance(data, (str, Path)):
+        with open(data, errors='ignore') as f:
+            data = yaml.safe_load(f)  # dictionary
+
+    # Parse yaml
+    path = extract_dir or Path(data.get('path') or '')  # optional 'path' default to '.'
+    for k in 'train', 'val', 'test':
+        if data.get(k):  # prepend path
+            data[k] = str(path / data[k]) if isinstance(data[k], str) else [str(path / x) for x in data[k]]
+
+    assert 'nc' in data, "Dataset 'nc' key missing."
+    if 'names' not in data:
+        data['names'] = [f'class{i}' for i in range(data['nc'])]  # assign class names if missing
+    train, val, test, s = [data.get(x) for x in ('train', 'val', 'test', 'download')]
+    if val:
+        val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])]  # val path
+        if not all(x.exists() for x in val):
+            if 'kp_bbox' in data.keys():
+                print('Writing dataset labels to {}...'.format(os.path.join(data['path'], data['labels'])))
+                write_kp_labels(data)
+            else:
+                print('\nWARNING: Dataset not found, nonexistent paths: %s' % [str(x) for x in val if not x.exists()])
+                if s and autodownload:  # download script
+                    if s.startswith('http') and s.endswith('.zip'):  # URL
+                        f = Path(s).name  # filename
+                        print(f'Downloading {s} ...')
+                        torch.hub.download_url_to_file(s, f)
+                        root = path.parent if 'path' in data else '..'  # unzip directory i.e. '../'
+                        Path(root).mkdir(parents=True, exist_ok=True)  # create root
+                        r = os.system(f'unzip -q {f} -d {root} && rm {f}')  # unzip
+                    elif s.startswith('bash '):  # bash script
+                        print(f'Running {s} ...')
+                        r = os.system(s)
+                    else:  # python script
+                        r = exec(s, {'yaml': data})  # return None
+                    print('Dataset autodownload %s\n' % ('success' if r in (0, None) else 'failure'))  # print result
+                else:
+                    raise Exception('Dataset not found.')
+
+    return data  # dictionary
+
+
+def download(url, dir='.', unzip=True, delete=True, curl=False, threads=1):
+    # Multi-threaded file download and unzip function, used in data.yaml for autodownload
+    def download_one(url, dir):
+        # Download 1 file
+        f = dir / Path(url).name  # filename
+        if Path(url).is_file():  # exists in current path
+            Path(url).rename(f)  # move to dir
+        elif not f.exists():
+            print(f'Downloading {url} to {f}...')
+            if curl:
+                os.system(f"curl -L '{url}' -o '{f}' --retry 9 -C -")  # curl download, retry and resume on fail
+            else:
+                torch.hub.download_url_to_file(url, f, progress=True)  # torch download
+        if unzip and f.suffix in ('.zip', '.gz'):
+            print(f'Unzipping {f}...')
+            if f.suffix == '.zip':
+                s = f'unzip -qo {f} -d {dir}'  # unzip -quiet -overwrite
+            elif f.suffix == '.gz':
+                s = f'tar xfz {f} --directory {f.parent}'  # unzip
+            if delete:  # delete zip file after unzip
+                s += f' && rm {f}'
+            os.system(s)
+
+    dir = Path(dir)
+    dir.mkdir(parents=True, exist_ok=True)  # make directory
+    if threads > 1:
+        pool = ThreadPool(threads)
+        pool.imap(lambda x: download_one(*x), zip(url, repeat(dir)))  # multi-threaded
+        pool.close()
+        pool.join()
+    else:
+        for u in [url] if isinstance(url, (str, Path)) else url:
+            download_one(u, dir)
+
+
+def make_divisible(x, divisor):
+    # Returns x evenly divisible by divisor
+    return math.ceil(x / divisor) * divisor
+
+
+def clean_str(s):
+    # Cleans a string by replacing special characters with underscore _
+    return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s)
+
+
+def one_cycle(y1=0.0, y2=1.0, steps=100):
+    # lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf
+    return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1
+
+
+def colorstr(*input):
+    # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e.  colorstr('blue', 'hello world')
+    *args, string = input if len(input) > 1 else ('blue', 'bold', input[0])  # color arguments, string
+    colors = {'black': '\033[30m',  # basic colors
+              'red': '\033[31m',
+              'green': '\033[32m',
+              'yellow': '\033[33m',
+              'blue': '\033[34m',
+              'magenta': '\033[35m',
+              'cyan': '\033[36m',
+              'white': '\033[37m',
+              'bright_black': '\033[90m',  # bright colors
+              'bright_red': '\033[91m',
+              'bright_green': '\033[92m',
+              'bright_yellow': '\033[93m',
+              'bright_blue': '\033[94m',
+              'bright_magenta': '\033[95m',
+              'bright_cyan': '\033[96m',
+              'bright_white': '\033[97m',
+              'end': '\033[0m',  # misc
+              'bold': '\033[1m',
+              'underline': '\033[4m'}
+    return ''.join(colors[x] for x in args) + f'{string}' + colors['end']
+
+
+def labels_to_class_weights(labels, nc=80):
+    # Get class weights (inverse frequency) from training labels
+    if labels[0] is None:  # no labels loaded
+        return torch.Tensor()
+
+    labels = np.concatenate(labels, 0)  # labels.shape = (866643, 5) for COCO
+    classes = labels[:, 0].astype(np.int)  # labels = [class xywh]
+    weights = np.bincount(classes, minlength=nc)  # occurrences per class
+
+    # Prepend gridpoint count (for uCE training)
+    # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum()  # gridpoints per image
+    # weights = np.hstack([gpi * len(labels)  - weights.sum() * 9, weights * 9]) ** 0.5  # prepend gridpoints to start
+
+    weights[weights == 0] = 1  # replace empty bins with 1
+    weights = 1 / weights  # number of targets per class
+    weights /= weights.sum()  # normalize
+    return torch.from_numpy(weights)
+
+
+def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)):
+    # Produces image weights based on class_weights and image contents
+    class_counts = np.array([np.bincount(x[:, 0].astype(np.int), minlength=nc) for x in labels])
+    image_weights = (class_weights.reshape(1, nc) * class_counts).sum(1)
+    # index = random.choices(range(n), weights=image_weights, k=1)  # weight image sample
+    return image_weights
+
+
+def coco80_to_coco91_class():  # converts 80-index (val2014) to 91-index (paper)
+    # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/
+    # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n')
+    # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n')
+    # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)]  # darknet to coco
+    # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)]  # coco to darknet
+    x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34,
+         35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+         64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90]
+    return x
+
+
+def xyxy2xywh(x):
+    # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = (x[:, 0] + x[:, 2]) / 2  # x center
+    y[:, 1] = (x[:, 1] + x[:, 3]) / 2  # y center
+    y[:, 2] = x[:, 2] - x[:, 0]  # width
+    y[:, 3] = x[:, 3] - x[:, 1]  # height
+    return y
+
+
+def xywh2xyxy(x):
+    # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+
+def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0):
+    # Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw  # top left x
+    y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh  # top left y
+    y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw  # bottom right x
+    y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh  # bottom right y
+
+    # Convert keypoints from [x, y, v] normalized to [x, y]
+    if y.shape[-1] > 4:
+        nl = y.shape[0]
+        kp = y[:, 4:].reshape(nl, -1, 3)
+        kp[..., 0] *= w
+        kp[..., 0] += padw
+        kp[..., 1] *= h
+        kp[..., 1] += padh
+        y[:, 4:] = kp.reshape(nl, -1)
+
+    return y
+
+
+def xyxy2xywhn(x, w=640, h=640, clip=False, eps=0.0):
+    # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] normalized where xy1=top-left, xy2=bottom-right
+    if clip:
+        clip_coords(x, (h - eps, w - eps))  # warning: inplace clip
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = ((x[:, 0] + x[:, 2]) / 2) / w  # x center
+    y[:, 1] = ((x[:, 1] + x[:, 3]) / 2) / h  # y center
+    y[:, 2] = (x[:, 2] - x[:, 0]) / w  # width
+    y[:, 3] = (x[:, 3] - x[:, 1]) / h  # height
+
+    # convert keypoints from [x, y, v] to [x, y, v] normalized
+    if y.shape[-1] > 4:
+        nl = y.shape[0]
+        kp = y[:, 4:].reshape(nl, -1, 3)
+        kp[..., 0] /= w
+        kp[..., 1] /= h
+        y[:, 4:] = kp.reshape(nl, -1)
+
+    return y
+
+
+def xyn2xy(x, w=640, h=640, padw=0, padh=0):
+    # Convert normalized segments into pixel segments, shape (n,2)
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = w * x[:, 0] + padw  # top left x
+    y[:, 1] = h * x[:, 1] + padh  # top left y
+    return y
+
+
+def segment2box(segment, width=640, height=640):
+    # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy)
+    x, y = segment.T  # segment xy
+    inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height)
+    x, y, = x[inside], y[inside]
+    return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4))  # xyxy
+
+
+def segments2boxes(segments):
+    # Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh)
+    boxes = []
+    for s in segments:
+        x, y = s.T  # segment xy
+        boxes.append([x.min(), y.min(), x.max(), y.max()])  # cls, xyxy
+    return xyxy2xywh(np.array(boxes))  # cls, xywh
+
+
+def resample_segments(segments, n=1000):
+    # Up-sample an (n,2) segment
+    for i, s in enumerate(segments):
+        x = np.linspace(0, len(s) - 1, n)
+        xp = np.arange(len(s))
+        segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T  # segment xy
+    return segments
+
+
+def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None):
+    # Rescale coords (xyxy) from img1_shape to img0_shape
+    if ratio_pad is None:  # calculate from img0_shape
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
+        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
+    else:
+        gain = ratio_pad[0][0]
+        pad = ratio_pad[1]
+
+    nl = coords.shape[0]
+    coords = coords.reshape((nl, -1, 2))
+    coords[..., 0] -= pad[0]
+    coords[..., 1] -= pad[1]
+    coords /= gain
+    coords = coords.reshape(nl, -1)
+    clip_coords(coords, img0_shape)
+    return coords
+
+
+def clip_coords(boxes, shape):
+    # Clip bounding xyxy bounding boxes to image shape (height, width)
+    if isinstance(boxes, torch.Tensor):  # faster individually
+        boxes[:, 0].clamp_(0, shape[1])  # x1
+        boxes[:, 1].clamp_(0, shape[0])  # y1
+        boxes[:, 2].clamp_(0, shape[1])  # x2
+        boxes[:, 3].clamp_(0, shape[0])  # y2
+    else:  # np.array (faster grouped)
+        boxes[:, [0, 2]] = boxes[:, [0, 2]].clip(0, shape[1])  # x1, x2
+        boxes[:, [1, 3]] = boxes[:, [1, 3]].clip(0, shape[0])  # y1, y2
+
+
+def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
+                        labels=(), max_det=300):
+    """Runs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         list of detections, on (n,6) tensor per image [xyxy, conf, cls]
+    """
+
+    nc = prediction.shape[2] - 5  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Checks
+    assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
+    assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
+
+    # Settings
+    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
+    max_nms = 30000  # maximum number of boxes into torchvision.ops.nms()
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            l = labels[xi]
+            v = torch.zeros((len(l), nc + 5), device=x.device)
+            v[:, :4] = l[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(l)), l[:, 0].long() + 5] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 5:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, cls)
+        if multi_label:
+            i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
+        else:  # best class only
+            conf, j = x[:, 5:].max(1, keepdim=True)
+            x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # Apply finite constraint
+        # if not torch.isfinite(x).all():
+        #     x = x[torch.isfinite(x).all(1)]
+
+        # Check shape
+        n = x.shape[0]  # number of boxes
+        if not n:  # no boxes
+            continue
+        elif n > max_nms:  # excess boxes
+            x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence
+
+        # Batched NMS
+        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        if i.shape[0] > max_det:  # limit detections
+            i = i[:max_det]
+        if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            print(f'WARNING: NMS time limit {time_limit}s exceeded')
+            break  # time limit exceeded
+
+    return output
+
+
+def non_max_suppression_kp(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, max_det=300, num_coords=34):
+    """Runs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         list of detections, on (n,6) tensor per image [xyxy, conf, cls, keypoints]
+    """
+
+    nc = prediction.shape[2] - 5 - num_coords  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Checks
+    assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
+    assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
+
+    # Settings
+    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
+    max_nms = 30000  # maximum number of boxes into torchvision.ops.nms()
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0, 6 + num_coords), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
+        x = x[xc[xi]]  # confidence
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 5:-num_coords] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, cls)
+        conf, j = x[:, 5:-num_coords].max(1, keepdim=True)
+        kp = x[:, -num_coords:]
+        x = torch.cat((box, conf, j.float(), kp), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # Apply finite constraint
+        # if not torch.isfinite(x).all():
+        #     x = x[torch.isfinite(x).all(1)]
+
+        # Check shape
+        n = x.shape[0]  # number of boxes
+        if not n:  # no boxes
+            continue
+        elif n > max_nms:  # excess boxes
+            x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence
+
+        # Batched NMS
+        c = x[:, 5:6] * max_wh  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        if i.shape[0] > max_det:  # limit detections
+            i = i[:max_det]
+        if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            print(f'WARNING: NMS time limit {time_limit}s exceeded')
+            break  # time limit exceeded
+
+    return output
+
+
+def strip_optimizer(f='best.pt', s=''):  # from utils.general import *; strip_optimizer()
+    # Strip optimizer from 'f' to finalize training, optionally save as 's'
+    x = torch.load(f, map_location=torch.device('cpu'))
+    if x.get('ema'):
+        x['model'] = x['ema']  # replace model with ema
+    for k in 'optimizer', 'training_results', 'wandb_id', 'ema', 'updates':  # keys
+        x[k] = None
+    x['epoch'] = -1
+    x['model'].half()  # to FP16
+    for p in x['model'].parameters():
+        p.requires_grad = False
+    torch.save(x, s or f)
+    mb = os.path.getsize(s or f) / 1E6  # filesize
+    print(f"Optimizer stripped from {f},{(' saved as %s,' % s) if s else ''} {mb:.1f}MB")
+
+
+def print_mutation(results, hyp, save_dir, bucket):
+    evolve_csv, results_csv, evolve_yaml = save_dir / 'evolve.csv', save_dir / 'results.csv', save_dir / 'hyp_evolve.yaml'
+    keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95',
+            'val/box_loss', 'val/obj_loss', 'val/cls_loss') + tuple(hyp.keys())  # [results + hyps]
+    keys = tuple(x.strip() for x in keys)
+    vals = results + tuple(hyp.values())
+    n = len(keys)
+
+    # Download (optional)
+    if bucket:
+        url = f'gs://{bucket}/evolve.csv'
+        if gsutil_getsize(url) > (os.path.getsize(evolve_csv) if os.path.exists(evolve_csv) else 0):
+            os.system(f'gsutil cp {url} {save_dir}')  # download evolve.csv if larger than local
+
+    # Log to evolve.csv
+    s = '' if evolve_csv.exists() else (('%20s,' * n % keys).rstrip(',') + '\n')  # add header
+    with open(evolve_csv, 'a') as f:
+        f.write(s + ('%20.5g,' * n % vals).rstrip(',') + '\n')
+
+    # Print to screen
+    print(colorstr('evolve: ') + ', '.join(f'{x.strip():>20s}' for x in keys))
+    print(colorstr('evolve: ') + ', '.join(f'{x:20.5g}' for x in vals), end='\n\n\n')
+
+    # Save yaml
+    with open(evolve_yaml, 'w') as f:
+        data = pd.read_csv(evolve_csv)
+        data = data.rename(columns=lambda x: x.strip())  # strip keys
+        i = np.argmax(fitness(data.values[:, :7]))  #
+        f.write(f'# YOLOv5 Hyperparameter Evolution Results\n' +
+                f'# Best generation: {i}\n' +
+                f'# Last generation: {len(data)}\n' +
+                f'# ' + ', '.join(f'{x.strip():>20s}' for x in keys[:7]) + '\n' +
+                f'# ' + ', '.join(f'{x:>20.5g}' for x in data.values[i, :7]) + '\n\n')
+        yaml.safe_dump(hyp, f, sort_keys=False)
+
+    if bucket:
+        os.system(f'gsutil cp {evolve_csv} {evolve_yaml} gs://{bucket}')  # upload
+
+
+def apply_classifier(x, model, img, im0):
+    # Apply a second stage classifier to yolo outputs
+    im0 = [im0] if isinstance(im0, np.ndarray) else im0
+    for i, d in enumerate(x):  # per image
+        if d is not None and len(d):
+            d = d.clone()
+
+            # Reshape and pad cutouts
+            b = xyxy2xywh(d[:, :4])  # boxes
+            b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1)  # rectangle to square
+            b[:, 2:] = b[:, 2:] * 1.3 + 30  # pad
+            d[:, :4] = xywh2xyxy(b).long()
+
+            # Rescale boxes from img_size to im0 size
+            scale_coords(img.shape[2:], d[:, :4], im0[i].shape)
+
+            # Classes
+            pred_cls1 = d[:, 5].long()
+            ims = []
+            for j, a in enumerate(d):  # per item
+                cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])]
+                im = cv2.resize(cutout, (224, 224))  # BGR
+                # cv2.imwrite('example%i.jpg' % j, cutout)
+
+                im = im[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
+                im = np.ascontiguousarray(im, dtype=np.float32)  # uint8 to float32
+                im /= 255.0  # 0 - 255 to 0.0 - 1.0
+                ims.append(im)
+
+            pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1)  # classifier prediction
+            x[i] = x[i][pred_cls1 == pred_cls2]  # retain matching class detections
+
+    return x
+
+
+def save_one_box(xyxy, im, file='image.jpg', gain=1.02, pad=10, square=False, BGR=False, save=True):
+    # Save image crop as {file} with crop size multiple {gain} and {pad} pixels. Save and/or return crop
+    xyxy = torch.tensor(xyxy).view(-1, 4)
+    b = xyxy2xywh(xyxy)  # boxes
+    if square:
+        b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1)  # attempt rectangle to square
+    b[:, 2:] = b[:, 2:] * gain + pad  # box wh * gain + pad
+    xyxy = xywh2xyxy(b).long()
+    clip_coords(xyxy, im.shape)
+    crop = im[int(xyxy[0, 1]):int(xyxy[0, 3]), int(xyxy[0, 0]):int(xyxy[0, 2]), ::(1 if BGR else -1)]
+    if save:
+        cv2.imwrite(str(increment_path(file, mkdir=True).with_suffix('.jpg')), crop)
+    return crop
+
+
+def increment_path(path, exist_ok=False, sep='', mkdir=False):
+    # Increment file or directory path, i.e. runs/exp --> runs/exp{sep}2, runs/exp{sep}3, ... etc.
+    path = Path(path)  # os-agnostic
+    if path.exists() and not exist_ok:
+        suffix = path.suffix
+        path = path.with_suffix('')
+        dirs = glob.glob(f"{path}{sep}*")  # similar paths
+        matches = [re.search(rf"%s{sep}(\d+)" % path.stem, d) for d in dirs]
+        i = [int(m.groups()[0]) for m in matches if m]  # indices
+        n = max(i) + 1 if i else 2  # increment number
+        path = Path(f"{path}{sep}{n}{suffix}")  # update path
+    dir = path if path.suffix == '' else path.parent  # directory
+    if not dir.exists() and mkdir:
+        dir.mkdir(parents=True, exist_ok=True)  # make directory
+    return path

utils/labels.py

@@ -0,0 +1,104 @@
+import os, os.path as osp
+import argparse
+import numpy as np
+import yaml
+from tqdm import tqdm
+
+
+def write_kp_labels(data):
+    assert not osp.isdir(osp.join(data['path'], data['labels'])), \
+        'Labels already generated. Remove or choose new name for labels.'
+
+    is_coco = 'coco' in data['path']
+    if is_coco:
+        from pycocotools.coco import COCO
+    else:
+        from crowdposetools.coco import COCO
+
+    splits = [osp.splitext(osp.split(data[s])[-1])[0] for s in ['train', 'val', 'test'] if s in data]
+    annotations = [osp.join(data['path'], data['{}_annotations'.format(s)]) for s in ['train', 'val', 'test'] if s in data]
+    test_split = [0 if s in ['train', 'val'] else 1 for s in ['train', 'val', 'test'] if s in data]
+    img_txt_dir = osp.join(data['path'], data['labels'], 'img_txt')
+    os.makedirs(img_txt_dir, exist_ok=True)
+
+    for split, annot, is_test in zip(splits, annotations, test_split):
+        img_txt_path = osp.join(img_txt_dir, '{}.txt'.format(split))
+        labels_path = osp.join(data['path'], '{}/{}'.format(data['labels'], split if is_coco else ''))
+        if not is_test:
+            os.makedirs(labels_path, exist_ok=True)
+        coco = COCO(annot)
+        if not is_test:
+            pbar = tqdm(coco.anns.keys(), total=len(coco.anns.keys()))
+            pbar.desc = 'Writing {} labels to {}'.format(split, labels_path)
+            for id in pbar:
+                a = coco.anns[id]
+
+                if a['image_id'] not in coco.imgs:
+                    continue
+
+                if 'train' in split:
+                    if is_coco and a['iscrowd']:
+                        continue
+
+                img_info = coco.imgs[a['image_id']]
+                img_h, img_w = img_info['height'], img_info['width']
+                x, y, w, h = a['bbox']
+                xc, yc = x + w / 2, y + h / 2
+                xc /= img_w
+                yc /= img_h
+                w /= img_w
+                h /= img_h
+
+                keypoints = np.array(a['keypoints']).reshape([-1, 3])
+
+                # some of crowdpose keypoints are just outside image so clip to image extents
+                if not is_coco:
+                    keypoints[:, 0] = np.clip(keypoints[:, 0], 0, img_w)
+                    keypoints[:, 1] = np.clip(keypoints[:, 1], 0, img_h)
+
+                with open(osp.join(labels_path, '{}.txt'.format(osp.splitext(img_info['file_name'])[0])), 'a') as f:
+                    # write person object
+                    s = '{} {:.6f} {:.6f} {:.6f} {:.6f}'.format(0, xc, yc, w, h)
+                    if data['pose_obj']:
+                        for i, (x, y, v) in enumerate(keypoints):
+                            s += ' {:.6f} {:.6f} {:.6f}'.format(x / img_w, y / img_h, v)
+                    s += '\n'
+                    f.write(s)
+
+                    # write keypoint objects
+                    for i, (x, y, v) in enumerate(keypoints):
+                        if v:
+                            if isinstance(data['kp_bbox'], list):
+                                kp_bbox = data['kp_bbox'][i]
+                            else:
+                                kp_bbox = data['kp_bbox']
+
+                            s = '{} {:.6f} {:.6f} {:.6f} {:.6f}'.format(
+                                i + 1, x / img_w, y / img_h,
+                                kp_bbox * max(img_h, img_w) / img_w,
+                                kp_bbox * max(img_h, img_w) / img_h)
+
+                            if data['pose_obj']:
+                                for _ in range(keypoints.shape[0]):
+                                    s += ' {:.6f} {:.6f} {:.6f}'.format(0, 0, 0)
+                            s += '\n'
+                            f.write(s)
+            pbar.close()
+
+        with open(img_txt_path, 'w') as f:
+            for img_info in coco.imgs.values():
+                f.write(osp.join(data['path'], 'images',
+                                 '{}'.format(split if is_coco else ''),
+                                 img_info['file_name']) + '\n')
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', default='data/coco-kp.yaml')
+    args = parser.parse_args()
+
+    assert osp.isfile(args.data), 'Data config file not found at {}'.format(args.data)
+
+    with open(args.data, 'rb') as f:
+        data = yaml.safe_load(f)
+    write_kp_labels(data)

utils/loggers/__init__.py

@@ -0,0 +1,149 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Logging utils
+"""
+
+import warnings
+from threading import Thread
+
+import torch
+from torch.utils.tensorboard import SummaryWriter
+
+from utils.general import colorstr, emojis
+from utils.loggers.wandb.wandb_utils import WandbLogger
+from utils.plots import plot_images, plot_results
+from utils.torch_utils import de_parallel
+
+LOGGERS = ('csv', 'tb', 'wandb')  # text-file, TensorBoard, Weights & Biases
+
+try:
+    import wandb
+
+    assert hasattr(wandb, '__version__')  # verify package import not local dir
+except (ImportError, AssertionError):
+    wandb = None
+
+
+class Loggers():
+    # YOLOv5 Loggers class
+    def __init__(self, save_dir=None, weights=None, opt=None, hyp=None, logger=None, include=LOGGERS):
+        self.save_dir = save_dir
+        self.weights = weights
+        self.opt = opt
+        self.hyp = hyp
+        self.logger = logger  # for printing results to console
+        self.include = include
+        self.keys = ['train/box_loss', 'train/obj_loss', 'train/cls_loss',  'train/kp_loss', # train loss
+                     'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95',  # metrics
+                     'val/box_loss', 'val/obj_loss', 'val/cls_loss', 'val/kp_loss', # val loss
+                     'x/lr0', 'x/lr1', 'x/lr2']  # params
+        for k in LOGGERS:
+            setattr(self, k, None)  # init empty logger dictionary
+        self.csv = True  # always log to csv
+
+        # Message
+        if not wandb:
+            prefix = colorstr('Weights & Biases: ')
+            s = f"{prefix}run 'pip install wandb' to automatically track and visualize YOLOv5 🚀 runs (RECOMMENDED)"
+            print(emojis(s))
+
+        # TensorBoard
+        s = self.save_dir
+        if 'tb' in self.include and not self.opt.evolve:
+            prefix = colorstr('TensorBoard: ')
+            self.logger.info(f"{prefix}Start with 'tensorboard --logdir {s.parent}', view at http://localhost:6006/")
+            self.tb = SummaryWriter(str(s))
+
+        # W&B
+        if wandb and 'wandb' in self.include:
+            wandb_artifact_resume = isinstance(self.opt.resume, str) and self.opt.resume.startswith('wandb-artifact://')
+            run_id = torch.load(self.weights).get('wandb_id') if self.opt.resume and not wandb_artifact_resume else None
+            self.opt.hyp = self.hyp  # add hyperparameters
+            self.wandb = WandbLogger(self.opt, run_id)
+        else:
+            self.wandb = None
+
+    def on_pretrain_routine_end(self):
+        # Callback runs on pre-train routine end
+        paths = self.save_dir.glob('*labels*.jpg')  # training labels
+        if self.wandb:
+            self.wandb.log({"Labels": [wandb.Image(str(x), caption=x.name) for x in paths]})
+
+    def on_train_batch_end(self, ni, model, imgs, targets, paths, plots, sync_bn):
+        # Callback runs on train batch end
+        if plots:
+            if ni == 0:
+                if not sync_bn:  # tb.add_graph() --sync known issue https://github.com/ultralytics/yolov5/issues/3754
+                    with warnings.catch_warnings():
+                        warnings.simplefilter('ignore')  # suppress jit trace warning
+                        self.tb.add_graph(torch.jit.trace(de_parallel(model), imgs[0:1], strict=False), [])
+            if ni < 3:
+                f = self.save_dir / f'train_batch{ni}.jpg'  # filename
+                Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start()
+            if self.wandb and ni == 10:
+                files = sorted(self.save_dir.glob('train*.jpg'))
+                self.wandb.log({'Mosaics': [wandb.Image(str(f), caption=f.name) for f in files if f.exists()]})
+
+    def on_train_epoch_end(self, epoch):
+        # Callback runs on train epoch end
+        if self.wandb:
+            self.wandb.current_epoch = epoch + 1
+
+    def on_val_image_end(self, pred, predn, path, names, im):
+        # Callback runs on val image end
+        if self.wandb:
+            self.wandb.val_one_image(pred, predn, path, names, im)
+
+    def on_val_end(self):
+        # Callback runs on val end
+        if self.wandb:
+            files = sorted(self.save_dir.glob('val*.jpg'))
+            self.wandb.log({"Validation": [wandb.Image(str(f), caption=f.name) for f in files]})
+
+    def on_fit_epoch_end(self, vals, epoch, best_fitness, fi):
+        # Callback runs at the end of each fit (train+val) epoch
+        x = {k: v for k, v in zip(self.keys, vals)}  # dict
+        if self.csv:
+            file = self.save_dir / 'results.csv'
+            n = len(x) + 1  # number of cols
+            s = '' if file.exists() else (('%20s,' * n % tuple(['epoch'] + self.keys)).rstrip(',') + '\n')  # add header
+            with open(file, 'a') as f:
+                f.write(s + ('%20.5g,' * n % tuple([epoch] + vals)).rstrip(',') + '\n')
+
+        if self.tb:
+            for k, v in x.items():
+                self.tb.add_scalar(k, v, epoch)
+
+        if self.wandb:
+            self.wandb.log(x)
+            self.wandb.end_epoch(best_result=best_fitness == fi)
+
+    def on_model_save(self, last, epoch, final_epoch, best_fitness, fi):
+        # Callback runs on model save event
+        if self.wandb:
+            if ((epoch + 1) % self.opt.save_period == 0 and not final_epoch) and self.opt.save_period != -1:
+                self.wandb.log_model(last.parent, self.opt, epoch, fi, best_model=best_fitness == fi)
+
+    def on_train_end(self, last, best, plots, epoch):
+        # Callback runs on training end
+        if plots:
+            plot_results(file=self.save_dir / 'results.csv')  # save results.png
+        files = ['results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]]
+        files = [(self.save_dir / f) for f in files if (self.save_dir / f).exists()]  # filter
+
+        if self.tb:
+            import cv2
+            for f in files:
+                self.tb.add_image(f.stem, cv2.imread(str(f))[..., ::-1], epoch, dataformats='HWC')
+
+        if self.wandb:
+            self.wandb.log({"Results": [wandb.Image(str(f), caption=f.name) for f in files]})
+            # Calling wandb.log. TODO: Refactor this into WandbLogger.log_model
+            if not self.opt.evolve:
+                wandb.log_artifact(str(best if best.exists() else last), type='model',
+                                   name='run_' + self.wandb.wandb_run.id + '_model',
+                                   aliases=['latest', 'best', 'stripped'])
+                self.wandb.finish_run()
+            else:
+                self.wandb.finish_run()
+                self.wandb = WandbLogger(self.opt)

utils/loggers/wandb/README.md

@@ -0,0 +1,140 @@
+📚 This guide explains how to use **Weights & Biases** (W&B) with YOLOv5 🚀.
+ * [About Weights & Biases](#about-weights-&-biases)
+ * [First-Time Setup](#first-time-setup)
+ * [Viewing runs](#viewing-runs)
+ * [Advanced Usage: Dataset Versioning and Evaluation](#advanced-usage)
+ * [Reports: Share your work with the world!](#reports)
+
+## About Weights & Biases
+Think of [W&B](https://wandb.ai/site?utm_campaign=repo_yolo_wandbtutorial) like GitHub for machine learning models. With a few lines of code, save everything you need to debug, compare and reproduce your models — architecture, hyperparameters, git commits, model weights, GPU usage, and even datasets and predictions.
+  
+ Used by top researchers including teams at OpenAI, Lyft, Github, and MILA, W&B is part of the new standard of best practices for machine learning. How W&B can help you optimize your machine learning workflows:
+ 
+ * [Debug](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Free-2) model performance in real time
+ * [GPU usage](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#System-4), visualized automatically
+ * [Custom charts](https://wandb.ai/wandb/customizable-charts/reports/Powerful-Custom-Charts-To-Debug-Model-Peformance--VmlldzoyNzY4ODI) for powerful, extensible visualization
+ * [Share insights](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Share-8) interactively with collaborators
+ * [Optimize hyperparameters](https://docs.wandb.com/sweeps) efficiently
+ * [Track](https://docs.wandb.com/artifacts) datasets, pipelines, and production models
+ 
+ ## First-Time Setup
+<details open>
+ <summary> Toggle Details </summary>
+When you first train, W&B will prompt you to create a new account and will generate an **API key** for you. If you are an existing user you can retrieve your key from https://wandb.ai/authorize. This key is used to tell W&B where to log your data. You only need to supply your key once, and then it is remembered on the same device.
+ 
+ W&B will create a cloud **project** (default is 'YOLOv5') for your training runs, and each new training run will be provided a unique run **name** within that project as project/name. You can also manually set your project and run name as:
+ 
+ ```shell
+ $ python train.py --project ... --name ...
+ ```
+ 
+ <img alt="" width="800" src="https://user-images.githubusercontent.com/26833433/98183367-4acbc600-1f08-11eb-9a23-7266a4192355.jpg">
+ </details>
+ 
+## Viewing Runs
+<details open>
+  <summary> Toggle Details </summary>
+ Run information streams from your environment to the W&B cloud console as you train. This allows you to monitor and even cancel runs in <b>realtime</b> . All important information is logged:
+ 
+ * Training & Validation losses
+ * Metrics: Precision, Recall, mAP@0.5, mAP@0.5:0.95
+ * Learning Rate over time
+ * A bounding box debugging panel, showing the training progress over time
+ * GPU: Type, **GPU Utilization**, power, temperature, **CUDA memory usage**
+ * System: Disk I/0, CPU utilization, RAM memory usage
+ * Your trained model as W&B Artifact
+ * Environment: OS and Python types, Git repository and state, **training command**
+ 
+ <img alt="" width="800" src="https://user-images.githubusercontent.com/26833433/98184457-bd3da580-1f0a-11eb-8461-95d908a71893.jpg">
+</details>
+
+## Advanced Usage
+You can leverage W&B artifacts and Tables integration to easily visualize and manage your datasets, models and training evaluations. Here are some quick examples to get you started.
+<details open>
+ <h3>1. Visualize and Version Datasets</h3>
+ Log, visualize, dynamically query, and understand your data with <a href='https://docs.wandb.ai/guides/data-vis/tables'>W&B Tables</a>. You can use the following command to log your dataset as a W&B Table. This will generate a <code>{dataset}_wandb.yaml</code> file which can be used to train from dataset artifact.
+ <details>
+  <summary> <b>Usage</b> </summary>
+   <b>Code</b> <code> $ python utils/logger/wandb/log_dataset.py --project ... --name ... --data .. </code>
+   
+ ![Screenshot (64)](https://user-images.githubusercontent.com/15766192/128486078-d8433890-98a3-4d12-8986-b6c0e3fc64b9.png)
+ </details>
+  
+ <h3> 2: Train and Log Evaluation simultaneousy </h3>
+   This is an extension of the previous section, but it'll also training after uploading the dataset. <b> This also evaluation Table</b>
+   Evaluation table compares your predictions and ground truths across the validation set for each epoch. It uses the references to the already uploaded datasets,
+   so no images will be uploaded from your system more than once.
+ <details>
+  <summary> <b>Usage</b> </summary>
+   <b>Code</b> <code> $ python utils/logger/wandb/log_dataset.py --data ..  --upload_data </code>
+   
+![Screenshot (72)](https://user-images.githubusercontent.com/15766192/128979739-4cf63aeb-a76f-483f-8861-1c0100b938a5.png)
+ </details>
+  
+ <h3> 3: Train using dataset artifact </h3>
+   When you upload a dataset as described in the first section, you get a new config file with an added `_wandb` to its name. This file contains the information that 
+   can be used to train a model directly from the dataset artifact. <b> This also logs evaluation </b>
+ <details>
+  <summary> <b>Usage</b> </summary>
+   <b>Code</b> <code> $ python utils/logger/wandb/log_dataset.py --data {data}_wandb.yaml </code>
+   
+![Screenshot (72)](https://user-images.githubusercontent.com/15766192/128979739-4cf63aeb-a76f-483f-8861-1c0100b938a5.png)
+ </details>
+  
+   <h3> 4: Save model checkpoints as artifacts </h3>
+  To enable saving and versioning checkpoints of your experiment, pass `--save_period n` with the base cammand, where `n` represents checkpoint interval. 
+  You can also log both the dataset and model checkpoints simultaneously. If not passed, only the final model will be logged
+
+ <details>
+  <summary> <b>Usage</b> </summary>
+   <b>Code</b> <code> $ python train.py --save_period 1 </code>
+   
+![Screenshot (68)](https://user-images.githubusercontent.com/15766192/128726138-ec6c1f60-639d-437d-b4ee-3acd9de47ef3.png)
+ </details>
+  
+</details>
+
+ <h3> 5: Resume runs from checkpoint artifacts. </h3>
+Any run can be resumed using artifacts if the <code>--resume</code> argument starts with <code>wandb-artifact://</code> prefix followed by the run path, i.e, <code>wandb-artifact://username/project/runid </code>. This doesn't require the model checkpoint to be present on the local system.
+
+ <details>
+  <summary> <b>Usage</b> </summary>
+   <b>Code</b> <code> $ python train.py --resume wandb-artifact://{run_path} </code>
+  
+![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png)
+ </details>
+ 
+  <h3> 6: Resume runs from dataset artifact & checkpoint artifacts. </h3>
+ <b> Local dataset or model checkpoints are not required. This can be used to resume runs directly on a different device </b>
+ The syntax is same as the previous section, but you'll need to lof both the dataset and model checkpoints as artifacts, i.e, set bot <code>--upload_dataset</code> or 
+ train from <code>_wandb.yaml</code> file and set <code>--save_period</code>
+
+ <details>
+  <summary> <b>Usage</b> </summary>
+   <b>Code</b> <code> $ python train.py --resume wandb-artifact://{run_path} </code>
+  
+![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png)
+ </details>
+  
+</details>
+
+
+
+ <h3> Reports </h3>
+ W&B Reports can be created from your saved runs for sharing online. Once a report is created you will receive a link you can use to publically share your results. Here is an example report created from the COCO128 tutorial trainings of all four YOLOv5 models ([link](https://wandb.ai/glenn-jocher/yolov5_tutorial/reports/YOLOv5-COCO128-Tutorial-Results--VmlldzozMDI5OTY)).
+ 
+ <img alt="" width="800" src="https://user-images.githubusercontent.com/26833433/98185222-794ba000-1f0c-11eb-850f-3e9c45ad6949.jpg">
+ 
+ ## Environments
+ YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled):
+ 
+ * **Google Colab and Kaggle** notebooks with free GPU: [![Open In Colab](https://camo.githubusercontent.com/84f0493939e0c4de4e6dbe113251b4bfb5353e57134ffd9fcab6b8714514d4d1/68747470733a2f2f636f6c61622e72657365617263682e676f6f676c652e636f6d2f6173736574732f636f6c61622d62616467652e737667)](https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb) [![Open In Kaggle](https://camo.githubusercontent.com/a08ca511178e691ace596a95d334f73cf4ce06e83a5c4a5169b8bb68cac27bef/68747470733a2f2f6b6167676c652e636f6d2f7374617469632f696d616765732f6f70656e2d696e2d6b6167676c652e737667)](https://www.kaggle.com/ultralytics/yolov5)
+ * **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)
+ * **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart)
+ * **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) [![Docker Pulls](https://camo.githubusercontent.com/280faedaf431e4c0c24fdb30ec00a66d627404e5c4c498210d3f014dd58c2c7e/68747470733a2f2f696d672e736869656c64732e696f2f646f636b65722f70756c6c732f756c7472616c79746963732f796f6c6f76353f6c6f676f3d646f636b6572)](https://hub.docker.com/r/ultralytics/yolov5)
+ 
+ ## Status
+ ![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg)
+ 
+ If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), validation ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on MacOS, Windows, and Ubuntu every 24 hours and on every commit.
+

utils/loggers/wandb/log_dataset.py

@@ -0,0 +1,23 @@
+import argparse
+
+from wandb_utils import WandbLogger
+
+WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
+
+
+def create_dataset_artifact(opt):
+    logger = WandbLogger(opt, None, job_type='Dataset Creation')  # TODO: return value unused
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path')
+    parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset')
+    parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project')
+    parser.add_argument('--entity', default=None, help='W&B entity')
+    parser.add_argument('--name', type=str, default='log dataset', help='name of W&B run')
+
+    opt = parser.parse_args()
+    opt.resume = False  # Explicitly disallow resume check for dataset upload job
+
+    create_dataset_artifact(opt)

utils/loggers/wandb/sweep.py

@@ -0,0 +1,33 @@
+import sys
+from pathlib import Path
+
+import wandb
+
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[3].as_posix())  # add utils/ to path
+
+from train import train, parse_opt
+from utils.general import increment_path
+from utils.torch_utils import select_device
+
+
+def sweep():
+    wandb.init()
+    # Get hyp dict from sweep agent
+    hyp_dict = vars(wandb.config).get("_items")
+
+    # Workaround: get necessary opt args
+    opt = parse_opt(known=True)
+    opt.batch_size = hyp_dict.get("batch_size")
+    opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok or opt.evolve))
+    opt.epochs = hyp_dict.get("epochs")
+    opt.nosave = True
+    opt.data = hyp_dict.get("data")
+    device = select_device(opt.device, batch_size=opt.batch_size)
+
+    # train
+    train(hyp_dict, opt, device)
+
+
+if __name__ == "__main__":
+    sweep()

utils/loggers/wandb/sweep.yaml

@@ -0,0 +1,143 @@
+# Hyperparameters for training
+# To set range- 
+# Provide min and max values as:
+#      parameter:
+#         
+#         min: scalar
+#         max: scalar
+# OR
+#
+# Set a specific list of search space-
+#     parameter: 
+#         values: [scalar1, scalar2, scalar3...]
+#         
+# You can use grid, bayesian and hyperopt search strategy 
+# For more info on configuring sweeps visit - https://docs.wandb.ai/guides/sweeps/configuration
+
+program: utils/loggers/wandb/sweep.py
+method: random
+metric:
+  name: metrics/mAP_0.5
+  goal: maximize
+
+parameters:
+  # hyperparameters: set either min, max range or values list
+  data:
+    value: "data/coco128.yaml"
+  batch_size:
+    values: [64]
+  epochs:
+    values: [10]
+
+  lr0:
+    distribution: uniform
+    min: 1e-5
+    max: 1e-1
+  lrf:
+    distribution: uniform
+    min: 0.01
+    max: 1.0
+  momentum:
+    distribution: uniform
+    min: 0.6
+    max: 0.98
+  weight_decay:
+    distribution: uniform
+    min: 0.0
+    max: 0.001
+  warmup_epochs:
+    distribution: uniform
+    min: 0.0
+    max: 5.0
+  warmup_momentum:
+    distribution: uniform
+    min: 0.0
+    max: 0.95
+  warmup_bias_lr:
+    distribution: uniform
+    min: 0.0
+    max: 0.2
+  box:
+    distribution: uniform
+    min: 0.02
+    max: 0.2
+  cls:
+    distribution: uniform
+    min: 0.2
+    max: 4.0
+  cls_pw:
+    distribution: uniform
+    min: 0.5
+    max: 2.0
+  obj:
+    distribution: uniform
+    min: 0.2
+    max: 4.0
+  obj_pw:
+    distribution: uniform
+    min: 0.5
+    max: 2.0
+  iou_t:
+    distribution: uniform
+    min: 0.1
+    max: 0.7
+  anchor_t:
+    distribution: uniform
+    min: 2.0
+    max: 8.0
+  fl_gamma:
+    distribution: uniform
+    min: 0.0
+    max: 0.1
+  hsv_h:
+    distribution: uniform
+    min: 0.0
+    max: 0.1
+  hsv_s:
+    distribution: uniform
+    min: 0.0
+    max: 0.9
+  hsv_v:
+    distribution: uniform
+    min: 0.0
+    max: 0.9
+  degrees:
+    distribution: uniform
+    min: 0.0
+    max: 45.0
+  translate:
+    distribution: uniform
+    min: 0.0
+    max: 0.9
+  scale:
+    distribution: uniform
+    min: 0.0
+    max: 0.9
+  shear:
+    distribution: uniform
+    min: 0.0
+    max: 10.0
+  perspective:
+    distribution: uniform
+    min: 0.0
+    max: 0.001
+  flipud:
+    distribution: uniform
+    min: 0.0
+    max: 1.0
+  fliplr:
+    distribution: uniform
+    min: 0.0
+    max: 1.0
+  mosaic:
+    distribution: uniform
+    min: 0.0
+    max: 1.0
+  mixup:
+    distribution: uniform
+    min: 0.0
+    max: 1.0
+  copy_paste:
+    distribution: uniform
+    min: 0.0
+    max: 1.0

utils/loggers/wandb/wandb_utils.py

@@ -0,0 +1,519 @@
+"""Utilities and tools for tracking runs with Weights & Biases."""
+
+import logging
+import os
+import sys
+from contextlib import contextmanager
+from pathlib import Path
+
+import yaml
+from tqdm import tqdm
+
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[3].as_posix())  # add yolov5/ to path
+
+from utils.datasets import LoadImagesAndLabels
+from utils.datasets import img2label_paths
+from utils.general import check_dataset, check_file
+
+try:
+    import wandb
+
+    assert hasattr(wandb, '__version__')  # verify package import not local dir
+except (ImportError, AssertionError):
+    wandb = None
+
+RANK = int(os.getenv('RANK', -1))
+WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
+
+
+def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX):
+    return from_string[len(prefix):]
+
+
+def check_wandb_config_file(data_config_file):
+    wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1))  # updated data.yaml path
+    if Path(wandb_config).is_file():
+        return wandb_config
+    return data_config_file
+
+
+def check_wandb_dataset(data_file):
+    is_wandb_artifact = False
+    if check_file(data_file) and data_file.endswith('.yaml'):
+        with open(data_file, errors='ignore') as f:
+            data_dict = yaml.safe_load(f)
+        is_wandb_artifact = (data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX) or
+                             data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX))
+    if is_wandb_artifact:
+        return data_dict
+    else:
+        return check_dataset(data_file)
+
+
+def get_run_info(run_path):
+    run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX))
+    run_id = run_path.stem
+    project = run_path.parent.stem
+    entity = run_path.parent.parent.stem
+    model_artifact_name = 'run_' + run_id + '_model'
+    return entity, project, run_id, model_artifact_name
+
+
+def check_wandb_resume(opt):
+    process_wandb_config_ddp_mode(opt) if RANK not in [-1, 0] else None
+    if isinstance(opt.resume, str):
+        if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+            if RANK not in [-1, 0]:  # For resuming DDP runs
+                entity, project, run_id, model_artifact_name = get_run_info(opt.resume)
+                api = wandb.Api()
+                artifact = api.artifact(entity + '/' + project + '/' + model_artifact_name + ':latest')
+                modeldir = artifact.download()
+                opt.weights = str(Path(modeldir) / "last.pt")
+            return True
+    return None
+
+
+def process_wandb_config_ddp_mode(opt):
+    with open(check_file(opt.data), errors='ignore') as f:
+        data_dict = yaml.safe_load(f)  # data dict
+    train_dir, val_dir = None, None
+    if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX):
+        api = wandb.Api()
+        train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias)
+        train_dir = train_artifact.download()
+        train_path = Path(train_dir) / 'data/images/'
+        data_dict['train'] = str(train_path)
+
+    if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX):
+        api = wandb.Api()
+        val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias)
+        val_dir = val_artifact.download()
+        val_path = Path(val_dir) / 'data/images/'
+        data_dict['val'] = str(val_path)
+    if train_dir or val_dir:
+        ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml')
+        with open(ddp_data_path, 'w') as f:
+            yaml.safe_dump(data_dict, f)
+        opt.data = ddp_data_path
+
+
+class WandbLogger():
+    """Log training runs, datasets, models, and predictions to Weights & Biases.
+
+    This logger sends information to W&B at wandb.ai. By default, this information
+    includes hyperparameters, system configuration and metrics, model metrics,
+    and basic data metrics and analyses.
+
+    By providing additional command line arguments to train.py, datasets,
+    models and predictions can also be logged.
+
+    For more on how this logger is used, see the Weights & Biases documentation:
+    https://docs.wandb.com/guides/integrations/yolov5
+    """
+
+    def __init__(self, opt, run_id=None, job_type='Training'):
+        """
+        - Initialize WandbLogger instance
+        - Upload dataset if opt.upload_dataset is True
+        - Setup trainig processes if job_type is 'Training'
+
+        arguments:
+        opt (namespace) -- Commandline arguments for this run
+        run_id (str) -- Run ID of W&B run to be resumed
+        job_type (str) -- To set the job_type for this run 
+
+       """
+        # Pre-training routine --
+        self.job_type = job_type
+        self.wandb, self.wandb_run = wandb, None if not wandb else wandb.run
+        self.val_artifact, self.train_artifact = None, None
+        self.train_artifact_path, self.val_artifact_path = None, None
+        self.result_artifact = None
+        self.val_table, self.result_table = None, None
+        self.bbox_media_panel_images = []
+        self.val_table_path_map = None
+        self.max_imgs_to_log = 16
+        self.wandb_artifact_data_dict = None
+        self.data_dict = None
+        # It's more elegant to stick to 1 wandb.init call, but useful config data is overwritten in the WandbLogger's wandb.init call
+        if isinstance(opt.resume, str):  # checks resume from artifact
+            if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+                entity, project, run_id, model_artifact_name = get_run_info(opt.resume)
+                model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name
+                assert wandb, 'install wandb to resume wandb runs'
+                # Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config
+                self.wandb_run = wandb.init(id=run_id,
+                                            project=project,
+                                            entity=entity,
+                                            resume='allow',
+                                            allow_val_change=True)
+                opt.resume = model_artifact_name
+        elif self.wandb:
+            self.wandb_run = wandb.init(config=opt,
+                                        resume="allow",
+                                        project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem,
+                                        entity=opt.entity,
+                                        name=opt.name if opt.name != 'exp' else None,
+                                        job_type=job_type,
+                                        id=run_id,
+                                        allow_val_change=True) if not wandb.run else wandb.run
+        if self.wandb_run:
+            if self.job_type == 'Training':
+                if opt.upload_dataset:
+                    if not opt.resume:
+                        self.wandb_artifact_data_dict = self.check_and_upload_dataset(opt)
+
+                if opt.resume:
+                    # resume from artifact
+                    if isinstance(opt.resume, str) and opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+                        self.data_dict = dict(self.wandb_run.config.data_dict)
+                    else:  # local resume
+                        self.data_dict = check_wandb_dataset(opt.data)
+                else:
+                    self.data_dict = check_wandb_dataset(opt.data)
+                    self.wandb_artifact_data_dict = self.wandb_artifact_data_dict or self.data_dict
+
+                    # write data_dict to config. useful for resuming from artifacts. Do this only when not resuming.
+                    self.wandb_run.config.update({'data_dict': self.wandb_artifact_data_dict},
+                                                 allow_val_change=True)
+                self.setup_training(opt)
+
+            if self.job_type == 'Dataset Creation':
+                self.data_dict = self.check_and_upload_dataset(opt)
+
+    def check_and_upload_dataset(self, opt):
+        """
+        Check if the dataset format is compatible and upload it as W&B artifact
+
+        arguments:
+        opt (namespace)-- Commandline arguments for current run
+
+        returns:
+        Updated dataset info dictionary where local dataset paths are replaced by WAND_ARFACT_PREFIX links.
+        """
+        assert wandb, 'Install wandb to upload dataset'
+        config_path = self.log_dataset_artifact(opt.data,
+                                                opt.single_cls,
+                                                'YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem)
+        print("Created dataset config file ", config_path)
+        with open(config_path, errors='ignore') as f:
+            wandb_data_dict = yaml.safe_load(f)
+        return wandb_data_dict
+
+    def setup_training(self, opt):
+        """
+        Setup the necessary processes for training YOLO models:
+          - Attempt to download model checkpoint and dataset artifacts if opt.resume stats with WANDB_ARTIFACT_PREFIX
+          - Update data_dict, to contain info of previous run if resumed and the paths of dataset artifact if downloaded
+          - Setup log_dict, initialize bbox_interval 
+
+        arguments:
+        opt (namespace) -- commandline arguments for this run
+
+        """
+        self.log_dict, self.current_epoch = {}, 0
+        self.bbox_interval = opt.bbox_interval
+        if isinstance(opt.resume, str):
+            modeldir, _ = self.download_model_artifact(opt)
+            if modeldir:
+                self.weights = Path(modeldir) / "last.pt"
+                config = self.wandb_run.config
+                opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp = str(
+                    self.weights), config.save_period, config.batch_size, config.bbox_interval, config.epochs, \
+                                                                                                       config.hyp
+        data_dict = self.data_dict
+        if self.val_artifact is None:  # If --upload_dataset is set, use the existing artifact, don't download
+            self.train_artifact_path, self.train_artifact = self.download_dataset_artifact(data_dict.get('train'),
+                                                                                           opt.artifact_alias)
+            self.val_artifact_path, self.val_artifact = self.download_dataset_artifact(data_dict.get('val'),
+                                                                                       opt.artifact_alias)
+
+        if self.train_artifact_path is not None:
+            train_path = Path(self.train_artifact_path) / 'data/images/'
+            data_dict['train'] = str(train_path)
+        if self.val_artifact_path is not None:
+            val_path = Path(self.val_artifact_path) / 'data/images/'
+            data_dict['val'] = str(val_path)
+
+        if self.val_artifact is not None:
+            self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
+            self.result_table = wandb.Table(["epoch", "id", "ground truth", "prediction", "avg_confidence"])
+            self.val_table = self.val_artifact.get("val")
+            if self.val_table_path_map is None:
+                self.map_val_table_path()
+        if opt.bbox_interval == -1:
+            self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1
+        train_from_artifact = self.train_artifact_path is not None and self.val_artifact_path is not None
+        # Update the the data_dict to point to local artifacts dir
+        if train_from_artifact:
+            self.data_dict = data_dict
+
+    def download_dataset_artifact(self, path, alias):
+        """
+        download the model checkpoint artifact if the path starts with WANDB_ARTIFACT_PREFIX
+
+        arguments:
+        path -- path of the dataset to be used for training
+        alias (str)-- alias of the artifact to be download/used for training
+
+        returns:
+        (str, wandb.Artifact) -- path of the downladed dataset and it's corresponding artifact object if dataset
+        is found otherwise returns (None, None)
+        """
+        if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX):
+            artifact_path = Path(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias)
+            dataset_artifact = wandb.use_artifact(artifact_path.as_posix().replace("\\", "/"))
+            assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'"
+            datadir = dataset_artifact.download()
+            return datadir, dataset_artifact
+        return None, None
+
+    def download_model_artifact(self, opt):
+        """
+        download the model checkpoint artifact if the resume path starts with WANDB_ARTIFACT_PREFIX
+
+        arguments:
+        opt (namespace) -- Commandline arguments for this run
+        """
+        if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+            model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest")
+            assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist'
+            modeldir = model_artifact.download()
+            epochs_trained = model_artifact.metadata.get('epochs_trained')
+            total_epochs = model_artifact.metadata.get('total_epochs')
+            is_finished = total_epochs is None
+            assert not is_finished, 'training is finished, can only resume incomplete runs.'
+            return modeldir, model_artifact
+        return None, None
+
+    def log_model(self, path, opt, epoch, fitness_score, best_model=False):
+        """
+        Log the model checkpoint as W&B artifact
+
+        arguments:
+        path (Path)   -- Path of directory containing the checkpoints
+        opt (namespace) -- Command line arguments for this run
+        epoch (int)  -- Current epoch number
+        fitness_score (float) -- fitness score for current epoch 
+        best_model (boolean) -- Boolean representing if the current checkpoint is the best yet.
+        """
+        model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', type='model', metadata={
+            'original_url': str(path),
+            'epochs_trained': epoch + 1,
+            'save period': opt.save_period,
+            'project': opt.project,
+            'total_epochs': opt.epochs,
+            'fitness_score': fitness_score
+        })
+        model_artifact.add_file(str(path / 'last.pt'), name='last.pt')
+        wandb.log_artifact(model_artifact,
+                           aliases=['latest', 'last', 'epoch ' + str(self.current_epoch), 'best' if best_model else ''])
+        print("Saving model artifact on epoch ", epoch + 1)
+
+    def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False):
+        """
+        Log the dataset as W&B artifact and return the new data file with W&B links
+
+        arguments:
+        data_file (str) -- the .yaml file with information about the dataset like - path, classes etc.
+        single_class (boolean)  -- train multi-class data as single-class
+        project (str) -- project name. Used to construct the artifact path
+        overwrite_config (boolean) -- overwrites the data.yaml file if set to true otherwise creates a new 
+        file with _wandb postfix. Eg -> data_wandb.yaml
+
+        returns:
+        the new .yaml file with artifact links. it can be used to start training directly from artifacts
+        """
+        self.data_dict = check_dataset(data_file)  # parse and check
+        data = dict(self.data_dict)
+        nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names'])
+        names = {k: v for k, v in enumerate(names)}  # to index dictionary
+        self.train_artifact = self.create_dataset_table(LoadImagesAndLabels(
+            data['train'], rect=True, batch_size=1), names, name='train') if data.get('train') else None
+        self.val_artifact = self.create_dataset_table(LoadImagesAndLabels(
+            data['val'], rect=True, batch_size=1), names, name='val') if data.get('val') else None
+        if data.get('train'):
+            data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train')
+        if data.get('val'):
+            data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val')
+        path = Path(data_file).stem
+        path = (path if overwrite_config else path + '_wandb') + '.yaml'  # updated data.yaml path
+        data.pop('download', None)
+        data.pop('path', None)
+        with open(path, 'w') as f:
+            yaml.safe_dump(data, f)
+
+        if self.job_type == 'Training':  # builds correct artifact pipeline graph
+            self.wandb_run.use_artifact(self.val_artifact)
+            self.wandb_run.use_artifact(self.train_artifact)
+            self.val_artifact.wait()
+            self.val_table = self.val_artifact.get('val')
+            self.map_val_table_path()
+        else:
+            self.wandb_run.log_artifact(self.train_artifact)
+            self.wandb_run.log_artifact(self.val_artifact)
+        return path
+
+    def map_val_table_path(self):
+        """
+        Map the validation dataset Table like name of file -> it's id in the W&B Table.
+        Useful for - referencing artifacts for evaluation.
+        """
+        self.val_table_path_map = {}
+        print("Mapping dataset")
+        for i, data in enumerate(tqdm(self.val_table.data)):
+            self.val_table_path_map[data[3]] = data[0]
+
+    def create_dataset_table(self, dataset, class_to_id, name='dataset'):
+        """
+        Create and return W&B artifact containing W&B Table of the dataset.
+
+        arguments:
+        dataset (LoadImagesAndLabels) -- instance of LoadImagesAndLabels class used to iterate over the data to build Table
+        class_to_id (dict(int, str)) -- hash map that maps class ids to labels
+        name (str) -- name of the artifact
+
+        returns:
+        dataset artifact to be logged or used
+        """
+        # TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging
+        artifact = wandb.Artifact(name=name, type="dataset")
+        img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None
+        img_files = tqdm(dataset.img_files) if not img_files else img_files
+        for img_file in img_files:
+            if Path(img_file).is_dir():
+                artifact.add_dir(img_file, name='data/images')
+                labels_path = 'labels'.join(dataset.path.rsplit('images', 1))
+                artifact.add_dir(labels_path, name='data/labels')
+            else:
+                artifact.add_file(img_file, name='data/images/' + Path(img_file).name)
+                label_file = Path(img2label_paths([img_file])[0])
+                artifact.add_file(str(label_file),
+                                  name='data/labels/' + label_file.name) if label_file.exists() else None
+        table = wandb.Table(columns=["id", "train_image", "Classes", "name"])
+        class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()])
+        for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)):
+            box_data, img_classes = [], {}
+            for cls, *xywh in labels[:, 1:].tolist():
+                cls = int(cls)
+                box_data.append({"position": {"middle": [xywh[0], xywh[1]], "width": xywh[2], "height": xywh[3]},
+                                 "class_id": cls,
+                                 "box_caption": "%s" % (class_to_id[cls])})
+                img_classes[cls] = class_to_id[cls]
+            boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}}  # inference-space
+            table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), list(img_classes.values()),
+                           Path(paths).name)
+        artifact.add(table, name)
+        return artifact
+
+    def log_training_progress(self, predn, path, names):
+        """
+        Build evaluation Table. Uses reference from validation dataset table.
+
+        arguments:
+        predn (list): list of predictions in the native space in the format - [xmin, ymin, xmax, ymax, confidence, class]
+        path (str): local path of the current evaluation image 
+        names (dict(int, str)): hash map that maps class ids to labels
+        """
+        class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()])
+        box_data = []
+        total_conf = 0
+        for *xyxy, conf, cls in predn.tolist():
+            if conf >= 0.25:
+                box_data.append(
+                    {"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]},
+                     "class_id": int(cls),
+                     "box_caption": "%s %.3f" % (names[cls], conf),
+                     "scores": {"class_score": conf},
+                     "domain": "pixel"})
+                total_conf = total_conf + conf
+        boxes = {"predictions": {"box_data": box_data, "class_labels": names}}  # inference-space
+        id = self.val_table_path_map[Path(path).name]
+        self.result_table.add_data(self.current_epoch,
+                                   id,
+                                   self.val_table.data[id][1],
+                                   wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set),
+                                   total_conf / max(1, len(box_data))
+                                   )
+
+    def val_one_image(self, pred, predn, path, names, im):
+        """
+        Log validation data for one image. updates the result Table if validation dataset is uploaded and log bbox media panel
+
+        arguments:
+        pred (list): list of scaled predictions in the format - [xmin, ymin, xmax, ymax, confidence, class]
+        predn (list): list of predictions in the native space - [xmin, ymin, xmax, ymax, confidence, class]
+        path (str): local path of the current evaluation image 
+        """
+        if self.val_table and self.result_table:  # Log Table if Val dataset is uploaded as artifact
+            self.log_training_progress(predn, path, names)
+
+        if len(self.bbox_media_panel_images) < self.max_imgs_to_log and self.current_epoch > 0:
+            if self.current_epoch % self.bbox_interval == 0:
+                box_data = [{"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]},
+                             "class_id": int(cls),
+                             "box_caption": "%s %.3f" % (names[cls], conf),
+                             "scores": {"class_score": conf},
+                             "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()]
+                boxes = {"predictions": {"box_data": box_data, "class_labels": names}}  # inference-space
+                self.bbox_media_panel_images.append(wandb.Image(im, boxes=boxes, caption=path.name))
+
+    def log(self, log_dict):
+        """
+        save the metrics to the logging dictionary
+
+        arguments:
+        log_dict (Dict) -- metrics/media to be logged in current step
+        """
+        if self.wandb_run:
+            for key, value in log_dict.items():
+                self.log_dict[key] = value
+
+    def end_epoch(self, best_result=False):
+        """
+        commit the log_dict, model artifacts and Tables to W&B and flush the log_dict.
+
+        arguments:
+        best_result (boolean): Boolean representing if the result of this evaluation is best or not
+        """
+        if self.wandb_run:
+            with all_logging_disabled():
+                if self.bbox_media_panel_images:
+                    self.log_dict["Bounding Box Debugger/Images"] = self.bbox_media_panel_images
+                wandb.log(self.log_dict)
+                self.log_dict = {}
+                self.bbox_media_panel_images = []
+            if self.result_artifact:
+                self.result_artifact.add(self.result_table, 'result')
+                wandb.log_artifact(self.result_artifact, aliases=['latest', 'last', 'epoch ' + str(self.current_epoch),
+                                                                  ('best' if best_result else '')])
+
+                wandb.log({"evaluation": self.result_table})
+                self.result_table = wandb.Table(["epoch", "id", "ground truth", "prediction", "avg_confidence"])
+                self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
+
+    def finish_run(self):
+        """
+        Log metrics if any and finish the current W&B run
+        """
+        if self.wandb_run:
+            if self.log_dict:
+                with all_logging_disabled():
+                    wandb.log(self.log_dict)
+            wandb.run.finish()
+
+
+@contextmanager
+def all_logging_disabled(highest_level=logging.CRITICAL):
+    """ source - https://gist.github.com/simon-weber/7853144
+    A context manager that will prevent any logging messages triggered during the body from being processed.
+    :param highest_level: the maximum logging level in use.
+      This would only need to be changed if a custom level greater than CRITICAL is defined.
+    """
+    previous_level = logging.root.manager.disable
+    logging.disable(highest_level)
+    try:
+        yield
+    finally:
+        logging.disable(previous_level)

utils/loss.py

@@ -0,0 +1,262 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Loss functions
+"""
+
+import torch
+import torch.nn as nn
+
+from utils.metrics import bbox_iou
+from utils.torch_utils import is_parallel
+
+
+def smooth_BCE(eps=0.1):  # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
+    # return positive, negative label smoothing BCE targets
+    return 1.0 - 0.5 * eps, 0.5 * eps
+
+
+class BCEBlurWithLogitsLoss(nn.Module):
+    # BCEwithLogitLoss() with reduced missing label effects.
+    def __init__(self, alpha=0.05):
+        super(BCEBlurWithLogitsLoss, self).__init__()
+        self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none')  # must be nn.BCEWithLogitsLoss()
+        self.alpha = alpha
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+        pred = torch.sigmoid(pred)  # prob from logits
+        dx = pred - true  # reduce only missing label effects
+        # dx = (pred - true).abs()  # reduce missing label and false label effects
+        alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
+        loss *= alpha_factor
+        return loss.mean()
+
+
+class FocalLoss(nn.Module):
+    # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
+    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
+        super(FocalLoss, self).__init__()
+        self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = loss_fcn.reduction
+        self.loss_fcn.reduction = 'none'  # required to apply FL to each element
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+        # p_t = torch.exp(-loss)
+        # loss *= self.alpha * (1.000001 - p_t) ** self.gamma  # non-zero power for gradient stability
+
+        # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
+        pred_prob = torch.sigmoid(pred)  # prob from logits
+        p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
+        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
+        modulating_factor = (1.0 - p_t) ** self.gamma
+        loss *= alpha_factor * modulating_factor
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:  # 'none'
+            return loss
+
+
+class QFocalLoss(nn.Module):
+    # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
+    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
+        super(QFocalLoss, self).__init__()
+        self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = loss_fcn.reduction
+        self.loss_fcn.reduction = 'none'  # required to apply FL to each element
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+
+        pred_prob = torch.sigmoid(pred)  # prob from logits
+        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
+        modulating_factor = torch.abs(true - pred_prob) ** self.gamma
+        loss *= alpha_factor * modulating_factor
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:  # 'none'
+            return loss
+
+
+class ComputeLoss:
+    # Compute losses
+    def __init__(self, model, autobalance=False, num_coords=0):
+        super(ComputeLoss, self).__init__()
+        self.sort_obj_iou = False
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02])  # P3-P7
+        self.ssi = list(det.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance
+
+        if self.autobalance:
+            self.loss_coeffs = model.module.loss_coeffs if is_parallel(model) else model.loss_coeffs[-1]
+
+        # for k in 'na', 'nc', 'nl', 'anchors':
+        #     setattr(self, k, getattr(det, k))
+        self.num_coords = num_coords
+        self.na = det.na
+        self.nc = det.nc
+        self.nl = det.nl
+        self.anchors = det.anchors
+
+    def __call__(self, p, targets):  # predictions, targets, model
+        device = targets.device
+        lcls = torch.zeros(1, device=device)
+        lbox = torch.zeros(1, device=device)
+        lobj = torch.zeros(1, device=device)
+        lkps = torch.zeros(1, device=device)  # keypoint loss
+        tcls, tbox, tkps, indices, anchors = self.build_targets(p, targets)  # targets
+
+        # Losses
+        for i, pi in enumerate(p):  # layer index, layer predictions
+            b, a, gj, gi = indices[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
+
+            n = b.shape[0]  # number of targets
+            if n:
+                ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets
+
+                # Regression
+                pxy = ps[:, :2].sigmoid() * 2. - 0.5
+                pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]  # range [0, 4] * anchor
+                pbox = torch.cat((pxy, pwh), 1)  # predicted box
+                iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Keypoints
+                if self.num_coords:
+                    tkp = tkps[i]
+                    vis = tkp[..., 2] > 0
+                    tkp_vis = tkp[vis]
+                    if len(tkp_vis):
+                        pkp = ps[:, 5 + self.nc:].reshape((-1, self.num_coords // 2, 2))
+                        pkp = (pkp.sigmoid() * 4. - 2.) * anchors[i][:, None, :]  # range [-2, 2] * anchor
+                        pkp_vis = pkp[vis]
+                        l2 = torch.linalg.norm(pkp_vis - tkp_vis[..., :2], dim=-1)
+                        lkps += torch.mean(l2)
+
+                # Objectness
+                score_iou = iou.detach().clamp(0).type(tobj.dtype)
+                if self.sort_obj_iou:
+                    sort_id = torch.argsort(score_iou)
+                    b, a, gj, gi, score_iou = b[sort_id], a[sort_id], gj[sort_id], gi[sort_id], score_iou[sort_id]
+                tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * score_iou  # iou ratio
+
+                # Classification
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(ps[:, 5:5 + self.nc], self.cn, device=device)  # targets
+                    t[range(n), tcls[i]] = self.cp
+                    lcls += self.BCEcls(ps[:, 5:5 + self.nc], t)  # BCE
+
+                # Append targets to text file
+                # with open('targets.txt', 'a') as file:
+                #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
+
+            obji = self.BCEobj(pi[..., 4], tobj)
+            lobj += obji * self.balance[i]  # obj loss
+
+        lbox *= self.hyp['box']
+        lobj *= self.hyp['obj']
+        lcls *= self.hyp['cls']
+        lkps *= self.hyp['kp']
+
+        if self.autobalance:
+            loss = (lbox + lobj + lcls) / (torch.exp(2 * self.loss_coeffs[0])) + self.loss_coeffs[0]
+            loss += lkps / (torch.exp(2 * self.loss_coeffs[1])) + self.loss_coeffs[1]
+        else:
+            loss = lbox + lobj + lcls + lkps
+
+        bs = tobj.shape[0]  # batch size
+
+        return loss * bs, torch.cat((lbox, lobj, lcls, lkps)).detach()
+
+    def build_targets(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h, x1,y1,v1, ..., x17,y17,v17)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        tcls, tbox, tkps, indices, anch = [], [], [], [], []
+        gain = torch.ones(7 + self.num_coords * 3 // 2, device=targets.device)  # normalized to gridspace gain
+        ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor([[0, 0],
+                            [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
+                            # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+                            ], device=targets.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors = self.anchors[i]
+            xy_gain = torch.tensor(p[i].shape)[[3, 2]]
+            gain[2:4] = xy_gain
+            gain[4:6] = xy_gain
+            for j in range(self.num_coords // 2):
+                kp_idx = 6 + j * 3
+                gain[kp_idx:kp_idx + 2] = xy_gain
+
+            # Match targets to anchors
+            t = targets * gain
+            if nt:
+                # Matches
+                r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1. < g) & (gxy > 1.)).T
+                l, m = ((gxi % 1. < g) & (gxi > 1.)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            b = t[:, 0].long()  # image
+            c = t[:, 1].long()  # class
+            gxy = t[:, 2:4]  # grid xy
+            gwh = t[:, 4:6]  # grid wh
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid xy indices
+
+            if self.num_coords:
+                kp = t[:, 6:-1].reshape(-1, self.num_coords // 2, 3)
+                kp[..., :2] -= gij[:, None, :]  # grid kp relative to grid box anchor
+                tkps.append(kp)
+
+            # Append
+            a = t[:, -1].long()  # anchor indices
+            indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices
+            tbox.append(torch.cat((gxy - gij, gwh), 1))  # box
+            anch.append(anchors[a])  # anchors
+            tcls.append(c)  # class
+
+        return tcls, tbox, tkps, indices, anch

utils/metrics.py

@@ -0,0 +1,333 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Model validation metrics
+"""
+
+import math
+import warnings
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+
+
+def fitness(x):
+    # Model fitness as a weighted combination of metrics
+    w = [0.0, 0.0, 0.1, 0.9]  # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
+    return (x[:, :4] * w).sum(1)
+
+
+def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=()):
+    """ Compute the average precision, given the recall and precision curves.
+    Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
+    # Arguments
+        tp:  True positives (nparray, nx1 or nx10).
+        conf:  Objectness value from 0-1 (nparray).
+        pred_cls:  Predicted object classes (nparray).
+        target_cls:  True object classes (nparray).
+        plot:  Plot precision-recall curve at mAP@0.5
+        save_dir:  Plot save directory
+    # Returns
+        The average precision as computed in py-faster-rcnn.
+    """
+
+    # Sort by objectness
+    i = np.argsort(-conf)
+    tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
+
+    # Find unique classes
+    unique_classes = np.unique(target_cls)
+    nc = unique_classes.shape[0]  # number of classes, number of detections
+
+    # Create Precision-Recall curve and compute AP for each class
+    px, py = np.linspace(0, 1, 1000), []  # for plotting
+    ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
+    for ci, c in enumerate(unique_classes):
+        i = pred_cls == c
+        n_l = (target_cls == c).sum()  # number of labels
+        n_p = i.sum()  # number of predictions
+
+        if n_p == 0 or n_l == 0:
+            continue
+        else:
+            # Accumulate FPs and TPs
+            fpc = (1 - tp[i]).cumsum(0)
+            tpc = tp[i].cumsum(0)
+
+            # Recall
+            recall = tpc / (n_l + 1e-16)  # recall curve
+            r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0)  # negative x, xp because xp decreases
+
+            # Precision
+            precision = tpc / (tpc + fpc)  # precision curve
+            p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1)  # p at pr_score
+
+            # AP from recall-precision curve
+            for j in range(tp.shape[1]):
+                ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
+                if plot and j == 0:
+                    py.append(np.interp(px, mrec, mpre))  # precision at mAP@0.5
+
+    # Compute F1 (harmonic mean of precision and recall)
+    f1 = 2 * p * r / (p + r + 1e-16)
+    if plot:
+        plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names)
+        plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1')
+        plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision')
+        plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall')
+
+    i = f1.mean(0).argmax()  # max F1 index
+    return p[:, i], r[:, i], ap, f1[:, i], unique_classes.astype('int32')
+
+
+def compute_ap(recall, precision):
+    """ Compute the average precision, given the recall and precision curves
+    # Arguments
+        recall:    The recall curve (list)
+        precision: The precision curve (list)
+    # Returns
+        Average precision, precision curve, recall curve
+    """
+
+    # Append sentinel values to beginning and end
+    mrec = np.concatenate(([0.0], recall, [1.0]))
+    mpre = np.concatenate(([1.0], precision, [0.0]))
+
+    # Compute the precision envelope
+    mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
+
+    # Integrate area under curve
+    method = 'interp'  # methods: 'continuous', 'interp'
+    if method == 'interp':
+        x = np.linspace(0, 1, 101)  # 101-point interp (COCO)
+        ap = np.trapz(np.interp(x, mrec, mpre), x)  # integrate
+    else:  # 'continuous'
+        i = np.where(mrec[1:] != mrec[:-1])[0]  # points where x axis (recall) changes
+        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])  # area under curve
+
+    return ap, mpre, mrec
+
+
+class ConfusionMatrix:
+    # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix
+    def __init__(self, nc, conf=0.25, iou_thres=0.45):
+        self.matrix = np.zeros((nc + 1, nc + 1))
+        self.nc = nc  # number of classes
+        self.conf = conf
+        self.iou_thres = iou_thres
+
+    def process_batch(self, detections, labels):
+        """
+        Return intersection-over-union (Jaccard index) of boxes.
+        Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+        Arguments:
+            detections (Array[N, 6]), x1, y1, x2, y2, conf, class
+            labels (Array[M, 5]), class, x1, y1, x2, y2
+        Returns:
+            None, updates confusion matrix accordingly
+        """
+        detections = detections[detections[:, 4] > self.conf]
+        gt_classes = labels[:, 0].int()
+        detection_classes = detections[:, 5].int()
+        iou = box_iou(labels[:, 1:], detections[:, :4])
+
+        x = torch.where(iou > self.iou_thres)
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+        else:
+            matches = np.zeros((0, 3))
+
+        n = matches.shape[0] > 0
+        m0, m1, _ = matches.transpose().astype(np.int16)
+        for i, gc in enumerate(gt_classes):
+            j = m0 == i
+            if n and sum(j) == 1:
+                self.matrix[detection_classes[m1[j]], gc] += 1  # correct
+            else:
+                self.matrix[self.nc, gc] += 1  # background FP
+
+        if n:
+            for i, dc in enumerate(detection_classes):
+                if not any(m1 == i):
+                    self.matrix[dc, self.nc] += 1  # background FN
+
+    def matrix(self):
+        return self.matrix
+
+    def plot(self, normalize=True, save_dir='', names=()):
+        try:
+            import seaborn as sn
+
+            array = self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1E-6) if normalize else 1)  # normalize columns
+            array[array < 0.005] = np.nan  # don't annotate (would appear as 0.00)
+
+            fig = plt.figure(figsize=(12, 9), tight_layout=True)
+            sn.set(font_scale=1.0 if self.nc < 50 else 0.8)  # for label size
+            labels = (0 < len(names) < 99) and len(names) == self.nc  # apply names to ticklabels
+            with warnings.catch_warnings():
+                warnings.simplefilter('ignore')  # suppress empty matrix RuntimeWarning: All-NaN slice encountered
+                sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True,
+                           xticklabels=names + ['background FP'] if labels else "auto",
+                           yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1))
+            fig.axes[0].set_xlabel('True')
+            fig.axes[0].set_ylabel('Predicted')
+            fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250)
+            plt.close()
+        except Exception as e:
+            print(f'WARNING: ConfusionMatrix plot failure: {e}')
+
+    def print(self):
+        for i in range(self.nc + 1):
+            print(' '.join(map(str, self.matrix[i])))
+
+
+def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
+    # Returns the IoU of box1 to box2. box1 is 4, box2 is nx4
+    box2 = box2.T
+
+    # Get the coordinates of bounding boxes
+    if x1y1x2y2:  # x1, y1, x2, y2 = box1
+        b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
+        b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
+    else:  # transform from xywh to xyxy
+        b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
+        b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
+        b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
+        b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2
+
+    # Intersection area
+    inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
+            (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
+
+    # Union Area
+    w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
+    w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
+    union = w1 * h1 + w2 * h2 - inter + eps
+
+    iou = inter / union
+    if GIoU or DIoU or CIoU:
+        cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1)  # convex (smallest enclosing box) width
+        ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1)  # convex height
+        if CIoU or DIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
+            c2 = cw ** 2 + ch ** 2 + eps  # convex diagonal squared
+            rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 +
+                    (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4  # center distance squared
+            if DIoU:
+                return iou - rho2 / c2  # DIoU
+            elif CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
+                v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
+                with torch.no_grad():
+                    alpha = v / (v - iou + (1 + eps))
+                return iou - (rho2 / c2 + v * alpha)  # CIoU
+        else:  # GIoU https://arxiv.org/pdf/1902.09630.pdf
+            c_area = cw * ch + eps  # convex area
+            return iou - (c_area - union) / c_area  # GIoU
+    else:
+        return iou  # IoU
+
+
+def box_iou(box1, box2):
+    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+    Arguments:
+        box1 (Tensor[N, 4])
+        box2 (Tensor[M, 4])
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+
+    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    return inter / (area1[:, None] + area2 - inter)  # iou = inter / (area1 + area2 - inter)
+
+
+def bbox_ioa(box1, box2, eps=1E-7):
+    """ Returns the intersection over box2 area given box1, box2. Boxes are x1y1x2y2
+    box1:       np.array of shape(4)
+    box2:       np.array of shape(nx4)
+    returns:    np.array of shape(n)
+    """
+
+    box2 = box2.transpose()
+
+    # Get the coordinates of bounding boxes
+    b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
+    b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
+
+    # Intersection area
+    inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \
+                 (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0)
+
+    # box2 area
+    box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + eps
+
+    # Intersection over box2 area
+    return inter_area / box2_area
+
+
+def wh_iou(wh1, wh2):
+    # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2
+    wh1 = wh1[:, None]  # [N,1,2]
+    wh2 = wh2[None]  # [1,M,2]
+    inter = torch.min(wh1, wh2).prod(2)  # [N,M]
+    return inter / (wh1.prod(2) + wh2.prod(2) - inter)  # iou = inter / (area1 + area2 - inter)
+
+
+# Plots ----------------------------------------------------------------------------------------------------------------
+
+def plot_pr_curve(px, py, ap, save_dir='pr_curve.png', names=()):
+    # Precision-recall curve
+    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
+    py = np.stack(py, axis=1)
+
+    if 0 < len(names) < 21:  # display per-class legend if < 21 classes
+        for i, y in enumerate(py.T):
+            ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}')  # plot(recall, precision)
+    else:
+        ax.plot(px, py, linewidth=1, color='grey')  # plot(recall, precision)
+
+    ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean())
+    ax.set_xlabel('Recall')
+    ax.set_ylabel('Precision')
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
+    fig.savefig(Path(save_dir), dpi=250)
+    plt.close()
+
+
+def plot_mc_curve(px, py, save_dir='mc_curve.png', names=(), xlabel='Confidence', ylabel='Metric'):
+    # Metric-confidence curve
+    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
+
+    if 0 < len(names) < 21:  # display per-class legend if < 21 classes
+        for i, y in enumerate(py):
+            ax.plot(px, y, linewidth=1, label=f'{names[i]}')  # plot(confidence, metric)
+    else:
+        ax.plot(px, py.T, linewidth=1, color='grey')  # plot(confidence, metric)
+
+    y = py.mean(0)
+    ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}')
+    ax.set_xlabel(xlabel)
+    ax.set_ylabel(ylabel)
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
+    fig.savefig(Path(save_dir), dpi=250)
+    plt.close()

utils/plots.py

@@ -0,0 +1,437 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Plotting utils
+"""
+
+import math
+from copy import copy
+from pathlib import Path
+
+import cv2
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import seaborn as sn
+import torch
+from PIL import Image, ImageDraw, ImageFont
+
+from utils.general import is_ascii, xyxy2xywh, xywh2xyxy
+from utils.metrics import fitness
+
+# Settings
+matplotlib.rc('font', **{'size': 11})
+matplotlib.use('Agg')  # for writing to files only
+
+FILE = Path(__file__).absolute()
+ROOT = FILE.parents[1]  # yolov5/ dir
+
+
+class Colors:
+    # Ultralytics color palette https://ultralytics.com/
+    def __init__(self):
+        # hex = matplotlib.colors.TABLEAU_COLORS.values()
+        hex = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB',
+               '2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7')
+        self.palette = [self.hex2rgb('#' + c) for c in hex]
+        self.n = len(self.palette)
+
+    def __call__(self, i, bgr=False):
+        c = self.palette[int(i) % self.n]
+        return (c[2], c[1], c[0]) if bgr else c
+
+    @staticmethod
+    def hex2rgb(h):  # rgb order (PIL)
+        return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
+
+
+colors = Colors()  # create instance for 'from utils.plots import colors'
+
+
+def check_font(font='Arial.ttf', size=10):
+    # Return a PIL TrueType Font, downloading to ROOT dir if necessary
+    font = Path(font)
+    font = font if font.exists() else (ROOT / font.name)
+    try:
+        return ImageFont.truetype(str(font) if font.exists() else font.name, size)
+    except Exception as e:  # download if missing
+        url = "https://ultralytics.com/assets/" + font.name
+        print(f'Downloading {url} to {font}...')
+        torch.hub.download_url_to_file(url, str(font))
+        return ImageFont.truetype(str(font), size)
+
+
+class Annotator:
+    check_font()  # download TTF if necessary
+
+    # YOLOv5 Annotator for train/val mosaics and jpgs and detect/hub inference annotations
+    def __init__(self, im, line_width=None, font_size=None, font='Arial.ttf', pil=True):
+        assert im.data.contiguous, 'Image not contiguous. Apply np.ascontiguousarray(im) to Annotator() input images.'
+        self.pil = pil
+        if self.pil:  # use PIL
+            self.im = im if isinstance(im, Image.Image) else Image.fromarray(im)
+            self.draw = ImageDraw.Draw(self.im)
+            self.font = check_font(font, size=font_size or max(round(sum(self.im.size) / 2 * 0.035), 12))
+            self.fh = self.font.getsize('a')[1] - 3  # font height
+        else:  # use cv2
+            self.im = im
+        self.lw = line_width or max(round(sum(im.shape) / 2 * 0.003), 2)  # line width
+
+    def box_label(self, box, label='', color=(128, 128, 128), txt_color=(255, 255, 255)):
+        # Add one xyxy box to image with label
+        if self.pil or not is_ascii(label):
+            self.draw.rectangle(box, width=self.lw, outline=color)  # box
+            if label:
+                w = self.font.getsize(label)[0]  # text width
+                self.draw.rectangle([box[0], box[1] - self.fh, box[0] + w + 1, box[1] + 1], fill=color)
+                self.draw.text((box[0], box[1]), label, fill=txt_color, font=self.font, anchor='ls')
+        else:  # cv2
+            c1, c2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3]))
+            cv2.rectangle(self.im, c1, c2, color, thickness=self.lw, lineType=cv2.LINE_AA)
+            if label:
+                tf = max(self.lw - 1, 1)  # font thickness
+                w, h = cv2.getTextSize(label, 0, fontScale=self.lw / 3, thickness=tf)[0]
+                c2 = c1[0] + w, c1[1] - h - 3
+                cv2.rectangle(self.im, c1, c2, color, -1, cv2.LINE_AA)  # filled
+                cv2.putText(self.im, label, (c1[0], c1[1] - 2), 0, self.lw / 3, txt_color, thickness=tf,
+                            lineType=cv2.LINE_AA)
+
+    def rectangle(self, xy, fill=None, outline=None, width=1):
+        # Add rectangle to image (PIL-only)
+        self.draw.rectangle(xy, fill, outline, width)
+
+    def text(self, xy, text, txt_color=(255, 255, 255)):
+        # Add text to image (PIL-only)
+        w, h = self.font.getsize(text)  # text width, height
+        self.draw.text((xy[0], xy[1] - h + 1), text, fill=txt_color, font=self.font)
+
+    def result(self):
+        # Return annotated image as array
+        return np.asarray(self.im)
+
+
+def hist2d(x, y, n=100):
+    # 2d histogram used in labels.png and evolve.png
+    xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n)
+    hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges))
+    xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1)
+    yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1)
+    return np.log(hist[xidx, yidx])
+
+
+def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5):
+    from scipy.signal import butter, filtfilt
+
+    # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy
+    def butter_lowpass(cutoff, fs, order):
+        nyq = 0.5 * fs
+        normal_cutoff = cutoff / nyq
+        return butter(order, normal_cutoff, btype='low', analog=False)
+
+    b, a = butter_lowpass(cutoff, fs, order=order)
+    return filtfilt(b, a, data)  # forward-backward filter
+
+
+def output_to_target(output):
+    # Convert model output to target format [batch_id, class_id, x, y, w, h, conf]
+    targets = []
+    for i, o in enumerate(output):
+        for *box, conf, cls in o.cpu().numpy():
+            targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf])
+    return np.array(targets)
+
+
+def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=1920, max_subplots=16):
+    # Plot image grid with labels
+    if isinstance(images, torch.Tensor):
+        images = images.cpu().float().numpy()
+    if isinstance(targets, torch.Tensor):
+        targets = targets.cpu().numpy()
+    if np.max(images[0]) <= 1:
+        images *= 255.0  # de-normalise (optional)
+    bs, _, h, w = images.shape  # batch size, _, height, width
+    bs = min(bs, max_subplots)  # limit plot images
+    ns = np.ceil(bs ** 0.5)  # number of subplots (square)
+
+    # Build Image
+    mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8)  # init
+    for i, im in enumerate(images):
+        if i == max_subplots:  # if last batch has fewer images than we expect
+            break
+        x, y = int(w * (i // ns)), int(h * (i % ns))  # block origin
+        im = im.transpose(1, 2, 0)
+        mosaic[y:y + h, x:x + w, :] = im
+
+    # Resize (optional)
+    scale = max_size / ns / max(h, w)
+    if scale < 1:
+        h = math.ceil(scale * h)
+        w = math.ceil(scale * w)
+        mosaic = cv2.resize(mosaic, tuple(int(x * ns) for x in (w, h)))
+
+    # Annotate
+    fs = int((h + w) * ns * 0.01)  # font size
+    annotator = Annotator(mosaic, line_width=round(fs / 10), font_size=fs)
+    for i in range(i + 1):
+        x, y = int(w * (i // ns)), int(h * (i % ns))  # block origin
+        annotator.rectangle([x, y, x + w, y + h], None, (255, 255, 255), width=2)  # borders
+        if paths:
+            annotator.text((x + 5, y + 5 + h), text=Path(paths[i]).name[:40], txt_color=(220, 220, 220))  # filenames
+        if len(targets) > 0:
+            ti = targets[targets[:, 0] == i]  # image targets
+            boxes = xywh2xyxy(ti[:, 2:6]).T
+            classes = ti[:, 1].astype('int')
+            labels = ti.shape[1] == 6 or ti.shape[1] > 7  # labels if no conf column or pose objects
+            conf = None if labels else ti[:, 6]  # check for confidence presence (label vs pred)
+
+            if boxes.shape[1]:
+                if boxes.max() <= 1.01:  # if normalized with tolerance 0.01
+                    boxes[[0, 2]] *= w  # scale to pixels
+                    boxes[[1, 3]] *= h
+                elif scale < 1:  # absolute coords need scale if image scales
+                    boxes *= scale
+            boxes[[0, 2]] += x
+            boxes[[1, 3]] += y
+            for j, box in enumerate(boxes.T.tolist()):
+                cls = classes[j]
+                color = colors(cls)
+                cls = names[cls] if names else cls
+                if labels or conf[j] > 0.25:  # 0.25 conf thresh
+                    label = f'{cls}' if labels else f'{cls} {conf[j]:.1f}'
+                    annotator.box_label(box, label, color=color)
+    annotator.im.save(fname)  # save
+
+
+def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''):
+    # Plot LR simulating training for full epochs
+    optimizer, scheduler = copy(optimizer), copy(scheduler)  # do not modify originals
+    y = []
+    for _ in range(epochs):
+        scheduler.step()
+        y.append(optimizer.param_groups[0]['lr'])
+    plt.plot(y, '.-', label='LR')
+    plt.xlabel('epoch')
+    plt.ylabel('LR')
+    plt.grid()
+    plt.xlim(0, epochs)
+    plt.ylim(0)
+    plt.savefig(Path(save_dir) / 'LR.png', dpi=200)
+    plt.close()
+
+
+def plot_val_txt():  # from utils.plots import *; plot_val()
+    # Plot val.txt histograms
+    x = np.loadtxt('val.txt', dtype=np.float32)
+    box = xyxy2xywh(x[:, :4])
+    cx, cy = box[:, 0], box[:, 1]
+
+    fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True)
+    ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0)
+    ax.set_aspect('equal')
+    plt.savefig('hist2d.png', dpi=300)
+
+    fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True)
+    ax[0].hist(cx, bins=600)
+    ax[1].hist(cy, bins=600)
+    plt.savefig('hist1d.png', dpi=200)
+
+
+def plot_targets_txt():  # from utils.plots import *; plot_targets_txt()
+    # Plot targets.txt histograms
+    x = np.loadtxt('targets.txt', dtype=np.float32).T
+    s = ['x targets', 'y targets', 'width targets', 'height targets']
+    fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)
+    ax = ax.ravel()
+    for i in range(4):
+        ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std()))
+        ax[i].legend()
+        ax[i].set_title(s[i])
+    plt.savefig('targets.jpg', dpi=200)
+
+
+def plot_study_txt(path='', x=None):  # from utils.plots import *; plot_study_txt()
+    # Plot study.txt generated by val.py
+    plot2 = False  # plot additional results
+    if plot2:
+        ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)[1].ravel()
+
+    fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True)
+    # for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolov5s6', 'yolov5m6', 'yolov5l6', 'yolov5x6']]:
+    for f in sorted(Path(path).glob('study*.txt')):
+        y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T
+        x = np.arange(y.shape[1]) if x is None else np.array(x)
+        if plot2:
+            s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_preprocess (ms/img)', 't_inference (ms/img)', 't_NMS (ms/img)']
+            for i in range(7):
+                ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8)
+                ax[i].set_title(s[i])
+
+        j = y[3].argmax() + 1
+        ax2.plot(y[5, 1:j], y[3, 1:j] * 1E2, '.-', linewidth=2, markersize=8,
+                 label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO'))
+
+    ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5],
+             'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet')
+
+    ax2.grid(alpha=0.2)
+    ax2.set_yticks(np.arange(20, 60, 5))
+    ax2.set_xlim(0, 57)
+    ax2.set_ylim(30, 55)
+    ax2.set_xlabel('GPU Speed (ms/img)')
+    ax2.set_ylabel('COCO AP val')
+    ax2.legend(loc='lower right')
+    plt.savefig(str(Path(path).name) + '.png', dpi=300)
+
+
+def plot_labels(labels, names=(), save_dir=Path('')):
+    # plot dataset labels
+    print('Plotting labels... ')
+    c, b = labels[:, 0], labels[:, 1:].transpose()  # classes, boxes
+    nc = int(c.max() + 1)  # number of classes
+    x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height'])
+
+    # seaborn correlogram
+    sn.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9))
+    plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200)
+    plt.close()
+
+    # matplotlib labels
+    matplotlib.use('svg')  # faster
+    ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel()
+    y = ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8)
+    # [y[2].patches[i].set_color([x / 255 for x in colors(i)]) for i in range(nc)]  # update colors bug #3195
+    ax[0].set_ylabel('instances')
+    if 0 < len(names) < 30:
+        ax[0].set_xticks(range(len(names)))
+        ax[0].set_xticklabels(names, rotation=90, fontsize=10)
+    else:
+        ax[0].set_xlabel('classes')
+    sn.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9)
+    sn.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9)
+
+    # rectangles
+    labels[:, 1:3] = 0.5  # center
+    labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000
+    img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255)
+    for cls, *box in labels[:1000]:
+        ImageDraw.Draw(img).rectangle(box, width=1, outline=colors(cls))  # plot
+    ax[1].imshow(img)
+    ax[1].axis('off')
+
+    for a in [0, 1, 2, 3]:
+        for s in ['top', 'right', 'left', 'bottom']:
+            ax[a].spines[s].set_visible(False)
+
+    plt.savefig(save_dir / 'labels.jpg', dpi=200)
+    matplotlib.use('Agg')
+    plt.close()
+
+
+def profile_idetection(start=0, stop=0, labels=(), save_dir=''):
+    # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection()
+    ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel()
+    s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS']
+    files = list(Path(save_dir).glob('frames*.txt'))
+    for fi, f in enumerate(files):
+        try:
+            results = np.loadtxt(f, ndmin=2).T[:, 90:-30]  # clip first and last rows
+            n = results.shape[1]  # number of rows
+            x = np.arange(start, min(stop, n) if stop else n)
+            results = results[:, x]
+            t = (results[0] - results[0].min())  # set t0=0s
+            results[0] = x
+            for i, a in enumerate(ax):
+                if i < len(results):
+                    label = labels[fi] if len(labels) else f.stem.replace('frames_', '')
+                    a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5)
+                    a.set_title(s[i])
+                    a.set_xlabel('time (s)')
+                    # if fi == len(files) - 1:
+                    #     a.set_ylim(bottom=0)
+                    for side in ['top', 'right']:
+                        a.spines[side].set_visible(False)
+                else:
+                    a.remove()
+        except Exception as e:
+            print('Warning: Plotting error for %s; %s' % (f, e))
+    ax[1].legend()
+    plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200)
+
+
+def plot_evolve(evolve_csv='path/to/evolve.csv'):  # from utils.plots import *; plot_evolve()
+    # Plot evolve.csv hyp evolution results
+    evolve_csv = Path(evolve_csv)
+    data = pd.read_csv(evolve_csv)
+    keys = [x.strip() for x in data.columns]
+    x = data.values
+    f = fitness(x)
+    j = np.argmax(f)  # max fitness index
+    plt.figure(figsize=(10, 12), tight_layout=True)
+    matplotlib.rc('font', **{'size': 8})
+    for i, k in enumerate(keys[7:]):
+        v = x[:, 7 + i]
+        mu = v[j]  # best single result
+        plt.subplot(6, 5, i + 1)
+        plt.scatter(v, f, c=hist2d(v, f, 20), cmap='viridis', alpha=.8, edgecolors='none')
+        plt.plot(mu, f.max(), 'k+', markersize=15)
+        plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9})  # limit to 40 characters
+        if i % 5 != 0:
+            plt.yticks([])
+        print('%15s: %.3g' % (k, mu))
+    f = evolve_csv.with_suffix('.png')  # filename
+    plt.savefig(f, dpi=200)
+    plt.close()
+    print(f'Saved {f}')
+
+
+def plot_results(file='path/to/results.csv', dir=''):
+    # Plot training results.csv. Usage: from utils.plots import *; plot_results('path/to/results.csv')
+    save_dir = Path(file).parent if file else Path(dir)
+    fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True)
+    ax = ax.ravel()
+    files = list(save_dir.glob('results*.csv'))
+    assert len(files), f'No results.csv files found in {save_dir.resolve()}, nothing to plot.'
+    for fi, f in enumerate(files):
+        try:
+            data = pd.read_csv(f)
+            s = [x.strip() for x in data.columns]
+            x = data.values[:, 0]
+            for i, j in enumerate([1, 2, 3, 4, 5, 8, 9, 10, 6, 7]):
+                y = data.values[:, j]
+                # y[y == 0] = np.nan  # don't show zero values
+                ax[i].plot(x, y, marker='.', label=f.stem, linewidth=2, markersize=8)
+                ax[i].set_title(s[j], fontsize=12)
+                # if j in [8, 9, 10]:  # share train and val loss y axes
+                #     ax[i].get_shared_y_axes().join(ax[i], ax[i - 5])
+        except Exception as e:
+            print(f'Warning: Plotting error for {f}: {e}')
+    ax[1].legend()
+    fig.savefig(save_dir / 'results.png', dpi=200)
+    plt.close()
+
+
+def feature_visualization(x, module_type, stage, n=32, save_dir=Path('runs/detect/exp')):
+    """
+    x:              Features to be visualized
+    module_type:    Module type
+    stage:          Module stage within model
+    n:              Maximum number of feature maps to plot
+    save_dir:       Directory to save results
+    """
+    if 'Detect' not in module_type:
+        batch, channels, height, width = x.shape  # batch, channels, height, width
+        if height > 1 and width > 1:
+            f = f"stage{stage}_{module_type.split('.')[-1]}_features.png"  # filename
+
+            blocks = torch.chunk(x[0].cpu(), channels, dim=0)  # select batch index 0, block by channels
+            n = min(n, channels)  # number of plots
+            fig, ax = plt.subplots(math.ceil(n / 8), 8, tight_layout=True)  # 8 rows x n/8 cols
+            ax = ax.ravel()
+            plt.subplots_adjust(wspace=0.05, hspace=0.05)
+            for i in range(n):
+                ax[i].imshow(blocks[i].squeeze())  # cmap='gray'
+                ax[i].axis('off')
+
+            print(f'Saving {save_dir / f}... ({n}/{channels})')
+            plt.savefig(save_dir / f, dpi=300, bbox_inches='tight')
+            plt.close()

utils/torch_utils.py

@@ -0,0 +1,350 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+PyTorch utils
+"""
+
+import datetime
+import logging
+import math
+import os
+import platform
+import subprocess
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+import torch.backends.cudnn as cudnn
+import torch.distributed as dist
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision
+
+try:
+    import thop  # for FLOPs computation
+except ImportError:
+    thop = None
+
+LOGGER = logging.getLogger(__name__)
+
+
+@contextmanager
+def torch_distributed_zero_first(local_rank: int):
+    """
+    Decorator to make all processes in distributed training wait for each local_master to do something.
+    """
+    if local_rank not in [-1, 0]:
+        dist.barrier(device_ids=[local_rank])
+    yield
+    if local_rank == 0:
+        dist.barrier(device_ids=[0])
+
+
+def init_torch_seeds(seed=0):
+    # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html
+    torch.manual_seed(seed)
+    if seed == 0:  # slower, more reproducible
+        cudnn.benchmark, cudnn.deterministic = False, True
+    else:  # faster, less reproducible
+        cudnn.benchmark, cudnn.deterministic = True, False
+
+
+def date_modified(path=__file__):
+    # return human-readable file modification date, i.e. '2021-3-26'
+    t = datetime.datetime.fromtimestamp(Path(path).stat().st_mtime)
+    return f'{t.year}-{t.month}-{t.day}'
+
+
+def git_describe(path=Path(__file__).parent):  # path must be a directory
+    # return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe
+    s = f'git -C {path} describe --tags --long --always'
+    try:
+        return subprocess.check_output(s, shell=True, stderr=subprocess.STDOUT).decode()[:-1]
+    except subprocess.CalledProcessError as e:
+        return ''  # not a git repository
+
+
+def select_device(device='', batch_size=None):
+    # device = 'cpu' or '0' or '0,1,2,3'
+    s = f'YOLOv5 🚀 {git_describe() or date_modified()} torch {torch.__version__} '  # string
+    device = str(device).strip().lower().replace('cuda:', '')  # to string, 'cuda:0' to '0'
+    cpu = device == 'cpu'
+    if cpu:
+        os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # force torch.cuda.is_available() = False
+    elif device:  # non-cpu device requested
+        os.environ['CUDA_VISIBLE_DEVICES'] = device  # set environment variable
+        assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested'  # check availability
+
+    cuda = not cpu and torch.cuda.is_available()
+    if cuda:
+        devices = device.split(',') if device else '0'  # range(torch.cuda.device_count())  # i.e. 0,1,6,7
+        n = len(devices)  # device count
+        if n > 1 and batch_size:  # check batch_size is divisible by device_count
+            assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}'
+        space = ' ' * (len(s) + 1)
+        for i, d in enumerate(devices):
+            p = torch.cuda.get_device_properties(i)
+            s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n"  # bytes to MB
+    else:
+        s += 'CPU\n'
+
+    LOGGER.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s)  # emoji-safe
+    return torch.device('cuda:0' if cuda else 'cpu')
+
+
+def time_sync():
+    # pytorch-accurate time
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+    return time.time()
+
+
+def profile(input, ops, n=10, device=None):
+    # YOLOv5 speed/memory/FLOPs profiler
+    #
+    # Usage:
+    #     input = torch.randn(16, 3, 640, 640)
+    #     m1 = lambda x: x * torch.sigmoid(x)
+    #     m2 = nn.SiLU()
+    #     profile(input, [m1, m2], n=100)  # profile over 100 iterations
+
+    results = []
+    logging.basicConfig(format="%(message)s", level=logging.INFO)
+    device = device or select_device()
+    print(f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}"
+          f"{'input':>24s}{'output':>24s}")
+
+    for x in input if isinstance(input, list) else [input]:
+        x = x.to(device)
+        x.requires_grad = True
+        for m in ops if isinstance(ops, list) else [ops]:
+            m = m.to(device) if hasattr(m, 'to') else m  # device
+            m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m
+            tf, tb, t = 0., 0., [0., 0., 0.]  # dt forward, backward
+            try:
+                flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # GFLOPs
+            except:
+                flops = 0
+
+            try:
+                for _ in range(n):
+                    t[0] = time_sync()
+                    y = m(x)
+                    t[1] = time_sync()
+                    try:
+                        _ = (sum([yi.sum() for yi in y]) if isinstance(y, list) else y).sum().backward()
+                        t[2] = time_sync()
+                    except Exception as e:  # no backward method
+                        print(e)
+                        t[2] = float('nan')
+                    tf += (t[1] - t[0]) * 1000 / n  # ms per op forward
+                    tb += (t[2] - t[1]) * 1000 / n  # ms per op backward
+                mem = torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0  # (GB)
+                s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list'
+                s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list'
+                p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0  # parameters
+                print(f'{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}')
+                results.append([p, flops, mem, tf, tb, s_in, s_out])
+            except Exception as e:
+                print(e)
+                results.append(None)
+            torch.cuda.empty_cache()
+    return results
+
+
+def is_parallel(model):
+    # Returns True if model is of type DP or DDP
+    return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)
+
+
+def de_parallel(model):
+    # De-parallelize a model: returns single-GPU model if model is of type DP or DDP
+    return model.module if is_parallel(model) else model
+
+
+def intersect_dicts(da, db, exclude=()):
+    # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values
+    return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape}
+
+
+def initialize_weights(model):
+    for m in model.modules():
+        t = type(m)
+        if t is nn.Conv2d:
+            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+        elif t is nn.BatchNorm2d:
+            m.eps = 1e-3
+            m.momentum = 0.03
+        elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
+            m.inplace = True
+
+
+def find_modules(model, mclass=nn.Conv2d):
+    # Finds layer indices matching module class 'mclass'
+    return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]
+
+
+def sparsity(model):
+    # Return global model sparsity
+    a, b = 0., 0.
+    for p in model.parameters():
+        a += p.numel()
+        b += (p == 0).sum()
+    return b / a
+
+
+def prune(model, amount=0.3):
+    # Prune model to requested global sparsity
+    import torch.nn.utils.prune as prune
+    print('Pruning model... ', end='')
+    for name, m in model.named_modules():
+        if isinstance(m, nn.Conv2d):
+            prune.l1_unstructured(m, name='weight', amount=amount)  # prune
+            prune.remove(m, 'weight')  # make permanent
+    print(' %.3g global sparsity' % sparsity(model))
+
+
+def fuse_conv_and_bn(conv, bn):
+    # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
+    fusedconv = nn.Conv2d(conv.in_channels,
+                          conv.out_channels,
+                          kernel_size=conv.kernel_size,
+                          stride=conv.stride,
+                          padding=conv.padding,
+                          groups=conv.groups,
+                          bias=True).requires_grad_(False).to(conv.weight.device)
+
+    # prepare filters
+    w_conv = conv.weight.clone().view(conv.out_channels, -1)
+    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
+    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))
+
+    # prepare spatial bias
+    b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
+    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
+    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
+
+    return fusedconv
+
+
+def model_info(model, verbose=False, img_size=640):
+    # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
+    n_p = sum(x.numel() for x in model.parameters())  # number parameters
+    n_g = sum(x.numel() for x in model.parameters() if x.requires_grad)  # number gradients
+    if verbose:
+        print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma'))
+        for i, (name, p) in enumerate(model.named_parameters()):
+            name = name.replace('module_list.', '')
+            print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
+                  (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
+
+    try:  # FLOPs
+        from thop import profile
+        stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32
+        img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device)  # input
+        flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2  # stride GFLOPs
+        img_size = img_size if isinstance(img_size, list) else [img_size, img_size]  # expand if int/float
+        fs = ', %.1f GFLOPs' % (flops * img_size[0] / stride * img_size[1] / stride)  # 640x640 GFLOPs
+    except (ImportError, Exception):
+        fs = ''
+
+    LOGGER.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
+
+
+def load_classifier(name='resnet101', n=2):
+    # Loads a pretrained model reshaped to n-class output
+    model = torchvision.models.__dict__[name](pretrained=True)
+
+    # ResNet model properties
+    # input_size = [3, 224, 224]
+    # input_space = 'RGB'
+    # input_range = [0, 1]
+    # mean = [0.485, 0.456, 0.406]
+    # std = [0.229, 0.224, 0.225]
+
+    # Reshape output to n classes
+    filters = model.fc.weight.shape[1]
+    model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True)
+    model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True)
+    model.fc.out_features = n
+    return model
+
+
+def scale_img(img, ratio=1.0, same_shape=False, gs=32):  # img(16,3,256,416)
+    # scales img(bs,3,y,x) by ratio constrained to gs-multiple
+    if ratio == 1.0:
+        return img
+    else:
+        h, w = img.shape[2:]
+        s = (int(h * ratio), int(w * ratio))  # new size
+        img = F.interpolate(img, size=s, mode='bilinear', align_corners=False)  # resize
+        if not same_shape:  # pad/crop img
+            h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)]
+        return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean
+
+
+def copy_attr(a, b, include=(), exclude=()):
+    # Copy attributes from b to a, options to only include [...] and to exclude [...]
+    for k, v in b.__dict__.items():
+        if (len(include) and k not in include) or k.startswith('_') or k in exclude:
+            continue
+        else:
+            setattr(a, k, v)
+
+
+class EarlyStopping:
+    # YOLOv5 simple early stopper
+    def __init__(self, patience=30):
+        self.best_fitness = 0.0  # i.e. mAP
+        self.best_epoch = 0
+        self.patience = patience or float('inf')  # epochs to wait after fitness stops improving to stop
+        self.possible_stop = False  # possible stop may occur next epoch
+
+    def __call__(self, epoch, fitness):
+        if fitness >= self.best_fitness:  # >= 0 to allow for early zero-fitness stage of training
+            self.best_epoch = epoch
+            self.best_fitness = fitness
+        delta = epoch - self.best_epoch  # epochs without improvement
+        self.possible_stop = delta >= (self.patience - 1)  # possible stop may occur next epoch
+        stop = delta >= self.patience  # stop training if patience exceeded
+        if stop:
+            LOGGER.info(f'EarlyStopping patience {self.patience} exceeded, stopping training.')
+        return stop
+
+
+class ModelEMA:
+    """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models
+    Keep a moving average of everything in the model state_dict (parameters and buffers).
+    This is intended to allow functionality like
+    https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
+    A smoothed version of the weights is necessary for some training schemes to perform well.
+    This class is sensitive where it is initialized in the sequence of model init,
+    GPU assignment and distributed training wrappers.
+    """
+
+    def __init__(self, model, decay=0.9999, updates=0):
+        # Create EMA
+        self.ema = deepcopy(model.module if is_parallel(model) else model).eval()  # FP32 EMA
+        # if next(model.parameters()).device.type != 'cpu':
+        #     self.ema.half()  # FP16 EMA
+        self.updates = updates  # number of EMA updates
+        self.decay = lambda x: decay * (1 - math.exp(-x / 2000))  # decay exponential ramp (to help early epochs)
+        for p in self.ema.parameters():
+            p.requires_grad_(False)
+
+    def update(self, model):
+        # Update EMA parameters
+        with torch.no_grad():
+            self.updates += 1
+            d = self.decay(self.updates)
+
+            msd = model.module.state_dict() if is_parallel(model) else model.state_dict()  # model state_dict
+            for k, v in self.ema.state_dict().items():
+                if v.dtype.is_floating_point:
+                    v *= d
+                    v += (1. - d) * msd[k].detach()
+
+    def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):
+        # Update EMA attributes
+        copy_attr(self.ema, model, include, exclude)

val.py

@@ -0,0 +1,336 @@
+import argparse
+import json
+import os, os.path as osp
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).absolute()
+sys.path.append(FILE.parents[0].as_posix())  # add kapao/ to path
+
+from models.experimental import attempt_load
+from utils.datasets import create_dataloader
+from utils.augmentations import letterbox
+from utils.general import check_dataset, check_file, check_img_size, \
+    non_max_suppression_kp, scale_coords, set_logging, colorstr
+from utils.torch_utils import select_device, time_sync
+import tempfile
+import cv2
+
+PAD_COLOR = (114 / 255, 114 / 255, 114 / 255)
+
+
+def run_nms(data, model_out):
+    if data['iou_thres'] == data['iou_thres_kp'] and data['conf_thres_kp'] >= data['conf_thres']:
+        # Combined NMS saves ~0.2 ms / image
+        dets = non_max_suppression_kp(model_out, data['conf_thres'], data['iou_thres'], num_coords=data['num_coords'])
+        person_dets = [d[d[:, 5] == 0] for d in dets]
+        kp_dets = [d[d[:, 4] >= data['conf_thres_kp']] for d in dets]
+        kp_dets = [d[d[:, 5] > 0] for d in kp_dets]
+    else:
+        person_dets = non_max_suppression_kp(model_out, data['conf_thres'], data['iou_thres'],
+                                             classes=[0],
+                                             num_coords=data['num_coords'])
+
+        kp_dets = non_max_suppression_kp(model_out, data['conf_thres_kp'], data['iou_thres_kp'],
+                                         classes=list(range(1, 1 + len(data['kp_flip']))),
+                                         num_coords=data['num_coords'])
+    return person_dets, kp_dets
+
+
+def post_process_batch(data, imgs, paths, shapes, person_dets, kp_dets,
+                       two_stage=False, pad=0, device='cpu', model=None, origins=None):
+
+    batch_bboxes, batch_poses, batch_scores, batch_ids = [], [], [], []
+    n_fused = np.zeros(data['num_coords'] // 2)
+
+    if origins is None:  # used only for two-stage inference so set to 0 if None
+        origins = [np.array([0, 0, 0]) for _ in range(len(person_dets))]
+
+    # process each image in batch
+    for si, (pd, kpd, origin) in enumerate(zip(person_dets, kp_dets, origins)):
+        nd = pd.shape[0]
+        nkp = kpd.shape[0]
+
+        if nd:
+            path, shape = Path(paths[si]) if len(paths) else '', shapes[si][0]
+            img_id = int(osp.splitext(osp.split(path)[-1])[0]) if path else si
+
+            # TWO-STAGE INFERENCE (EXPERIMENTAL)
+            if two_stage:
+                gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+                crops, origins, crop_shapes = [], [], []
+
+                for bbox in pd[:, :4].cpu().numpy():
+                    x1, y1, x2, y2 = map(int, map(round, bbox))
+                    x1, x2 = max(x1, 0), min(x2, data['imgsz'])
+                    y1, y2 = max(y1, 0), min(y2, data['imgsz'])
+                    h0, w0 = y2 - y1, x2 - x1
+                    crop_shapes.append([(h0, w0)])
+                    crop = np.transpose(imgs[si][:, y1:y2, x1:x2].cpu().numpy(), (1, 2, 0))
+                    crop = cv2.copyMakeBorder(crop, pad, pad, pad, pad, cv2.BORDER_CONSTANT, value=PAD_COLOR)  # add padding
+                    h0 += 2 * pad
+                    w0 += 2 * pad
+                    origins = [np.array([x1 - pad, y1 - pad, 0])]
+                    crop_pre = letterbox(crop, data['imgsz'], color=PAD_COLOR, stride=gs, auto=False)[0]
+                    crop_input = torch.Tensor(np.transpose(np.expand_dims(crop_pre, axis=0), (0, 3, 1, 2))).to(device)
+
+                    out = model(crop_input, augment=True, kp_flip=data['kp_flip'], scales=data['scales'], flips=data['flips'])[0]
+                    person_dets, kp_dets = run_nms(data, out)
+                    _, poses, scores, img_ids, _ = post_process_batch(
+                        data, crop_input, paths, [[(h0, w0)]], person_dets, kp_dets, device=device, origins=origins)
+
+                    # map back to original image
+                    if len(poses):
+                        poses = np.stack(poses, axis=0)
+                        poses = poses[:, :, :2].reshape(poses.shape[0], -1)
+                        poses = scale_coords(imgs[si].shape[1:], poses, shape)
+                        poses = poses.reshape(poses.shape[0], data['num_coords'] // 2, 2)
+                        poses = np.concatenate((poses, np.zeros((poses.shape[0], data['num_coords'] // 2, 1))), axis=-1)
+                    poses = [p for p in poses]  # convert back to list
+
+            # SINGLE-STAGE INFERENCE
+            else:
+                scores = pd[:, 4].cpu().numpy()  # person detection score
+                bboxes = scale_coords(imgs[si].shape[1:], pd[:, :4], shape).round().cpu().numpy()
+                poses = scale_coords(imgs[si].shape[1:], pd[:, -data['num_coords']:], shape).cpu().numpy()
+                poses = poses.reshape((nd, -data['num_coords'], 2))
+                poses = np.concatenate((poses, np.zeros((nd, poses.shape[1], 1))), axis=-1)
+
+                if data['use_kp_dets'] and nkp:
+                    mask = scores > data['conf_thres_kp_person']
+                    poses_mask = poses[mask]
+
+                    if len(poses_mask):
+                        kpd[:, :4] = scale_coords(imgs[si].shape[1:], kpd[:, :4], shape)
+                        kpd = kpd[:, :6].cpu()
+
+                        for x1, y1, x2, y2, conf, cls in kpd:
+                            x, y = np.mean((x1, x2)), np.mean((y1, y2))
+                            pose_kps = poses_mask[:, int(cls - 1)]
+                            dist = np.linalg.norm(pose_kps[:, :2] - np.array([[x, y]]), axis=-1)
+                            kp_match = np.argmin(dist)
+                            if conf > pose_kps[kp_match, 2] and dist[kp_match] < data['overwrite_tol']:
+                                pose_kps[kp_match] = [x, y, conf]
+                                n_fused[int(cls - 1)] += 1
+                        poses[mask] = poses_mask
+
+                poses = [p + origin for p in poses]
+
+            batch_bboxes.extend(bboxes)
+            batch_poses.extend(poses)
+            batch_scores.extend(scores)
+            batch_ids.extend([img_id] * len(scores))
+
+    return batch_bboxes, batch_poses, batch_scores, batch_ids, n_fused
+
+
+@torch.no_grad()
+def run(data,
+        weights=None,  # model.pt path(s)
+        batch_size=16,  # batch size
+        imgsz=1280,  # inference size (pixels)
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.65,  # NMS IoU threshold
+        no_kp_dets=False,
+        conf_thres_kp=0.2,
+        iou_thres_kp=0.25,
+        conf_thres_kp_person=0.3,
+        overwrite_tol=50,  # pixels for kp det overwrite
+        scales=[1],
+        flips=[None],
+        rect=False,
+        half=True,  # use FP16 half-precision inference
+        model=None,
+        dataloader=None,
+        compute_loss=None,
+        two_stage=False,
+        pad=0
+        ):
+
+    if two_stage:
+        assert batch_size == 1, 'Batch size must be set to 1 for two-stage processing'
+        assert conf_thres >= 0.01, 'Confidence threshold must be >= 0.01 for two-stage processing'
+        assert not rect, 'Cannot use rectangular inference with two-stage processing'
+        assert not half, 'Two-stage processing must use full precision'
+
+    use_kp_dets = not no_kp_dets
+
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device = next(model.parameters()).device  # get model device
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Load model
+        model = attempt_load(weights, map_location=device)  # load FP32 model
+        gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+        imgsz = check_img_size(imgsz, s=gs)  # check image size
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # add inference settings to data dict
+    data['imgsz'] = imgsz
+    data['conf_thres'] = conf_thres
+    data['iou_thres'] = iou_thres
+    data['use_kp_dets'] = use_kp_dets
+    data['conf_thres_kp'] = conf_thres_kp
+    data['iou_thres_kp'] = iou_thres_kp
+    data['conf_thres_kp_person'] = conf_thres_kp_person
+    data['overwrite_tol'] = overwrite_tol
+    data['scales'] = scales
+    data['flips'] = flips
+
+    is_coco = 'coco' in data['path']
+    if is_coco:
+        from pycocotools.coco import COCO
+        from pycocotools.cocoeval import COCOeval
+    else:
+        from crowdposetools.coco import COCO
+        from crowdposetools.cocoeval import COCOeval
+
+    # Half
+    half &= device.type != 'cpu'  # half precision only supported on CUDA
+    if half:
+        model.half()
+
+    model.eval()
+    nc = int(data['nc'])  # number of classes
+
+    # Dataloader
+    if not training:
+        if device.type != 'cpu':
+            model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters())))  # run once
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task], data['labels'], imgsz, batch_size, gs, rect=rect,
+                                       prefix=colorstr(f'{task}: '), kp_flip=data['kp_flip'])[0]
+
+    seen = 0
+    mp, mr, map50, mAP, t0, t1, t2 = 0., 0., 0., 0., 0., 0., 0.
+    loss = torch.zeros(4, device=device)
+    json_dump = []
+    n_fused = np.zeros(data['num_coords'] // 2)
+
+    for batch_i, (imgs, targets, paths, shapes) in enumerate(tqdm(dataloader, desc='Processing {} images'.format(task))):
+        t_ = time_sync()
+        imgs = imgs.to(device, non_blocking=True)
+        imgs = imgs.half() if half else imgs.float()  # uint8 to fp16/32
+        imgs /= 255.0  # 0 - 255 to 0.0 - 1.0
+        targets = targets.to(device)
+        nb, _, height, width = imgs.shape  # batch size, channels, height, width
+        t = time_sync()
+        t0 += t - t_
+
+        # Run model
+        out, train_out = model(imgs, augment=True, kp_flip=data['kp_flip'], scales=data['scales'], flips=data['flips'])
+        t1 += time_sync() - t
+
+        # Compute loss
+        if train_out:  # only computed if no scale / flipping
+            if compute_loss:
+                loss += compute_loss([x.float() for x in train_out], targets)[1]  # box, obj, cls, kp
+
+        t = time_sync()
+
+        # NMS
+        person_dets, kp_dets = run_nms(data, out)
+
+        # Fuse keypoint and pose detections
+        _, poses, scores, img_ids, n_fused_batch = post_process_batch(
+            data, imgs, paths, shapes, person_dets, kp_dets, two_stage, pad, device, model)
+
+        t2 += time_sync() - t
+        seen += len(imgs)
+        n_fused += n_fused_batch
+
+        for i, (pose, score, img_id) in enumerate(zip(poses, scores, img_ids)):
+            json_dump.append({
+                'image_id': img_id,
+                'category_id': 1,
+                'keypoints': pose.reshape(-1).tolist(),
+                'score': float(score)  # person score
+            })
+
+    if not training:  # save json
+        save_dir, weights_name = osp.split(weights)
+        json_name = '{}_{}_c{}_i{}_ck{}_ik{}_ckp{}_t{}.json'.format(
+            task, osp.splitext(weights_name)[0],
+            conf_thres, iou_thres, conf_thres_kp, iou_thres_kp,
+            conf_thres_kp_person, overwrite_tol
+        )
+        json_path = osp.join(save_dir, json_name)
+    else:
+        tmp = tempfile.NamedTemporaryFile(mode='w+b')
+        json_path = tmp.name
+
+    with open(json_path, 'w') as f:
+        json.dump(json_dump, f)
+
+    if task in ('train', 'val'):
+        annot = osp.join(data['path'], data['{}_annotations'.format(task)])
+        coco = COCO(annot)
+        result = coco.loadRes(json_path)
+        eval = COCOeval(coco, result, iouType='keypoints')
+        eval.evaluate()
+        eval.accumulate()
+        eval.summarize()
+        mAP, map50 = eval.stats[:2]
+
+    if training:
+        tmp.close()
+
+    # Print speeds
+    t = tuple(x / seen * 1E3 for x in (t0, t1, t2))  # speeds per image
+    if not training and task != 'test':
+        os.rename(json_path, osp.splitext(json_path)[0] + '_ap{:.4f}.json'.format(mAP))
+        shape = (batch_size, 3, imgsz, imgsz)
+        print(f'Speed: %.3fms pre-process, %.3fms inference, %.3fms NMS per image at shape {shape}' % t)
+        print('Keypoint Objects Fused:', n_fused)
+    model.float()  # for training
+    return (mp, mr, map50, mAP, *(loss.cpu() / len(dataloader)).tolist()), np.zeros(nc), t  # for compatibility with train
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser(prog='val.py')
+    parser.add_argument('--data', type=str, default='data/coco-kp.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', default='kapao_s_coco.pt')
+    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
+    parser.add_argument('--imgsz', type=int, default=1280, help='inference size (pixels)')
+    parser.add_argument('--task', default='val', help='train, val, test')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.65, help='NMS IoU threshold')
+    parser.add_argument('--no-kp-dets', action='store_true', help='do not use keypoint objects')
+    parser.add_argument('--conf-thres-kp', type=float, default=0.2)
+    parser.add_argument('--conf-thres-kp-person', type=float, default=0.3)
+    parser.add_argument('--iou-thres-kp', type=float, default=0.25)
+    parser.add_argument('--overwrite-tol', type=int, default=50)
+    parser.add_argument('--scales', type=float, nargs='+', default=[1])
+    parser.add_argument('--flips', type=int, nargs='+', default=[-1])
+    parser.add_argument('--rect', action='store_true', help='rectangular input image')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--two-stage', action='store_true', help='use two-stage inference (experimental)')
+    parser.add_argument('--pad', type=int, default=0, help='padding for two-stage inference')
+    opt = parser.parse_args()
+    opt.flips = [None if f == -1 else f for f in opt.flips]
+    opt.data = check_file(opt.data)  # check file
+    return opt
+
+
+def main(opt):
+    set_logging()
+    print(colorstr('val: ') + ', '.join(f'{k}={v}' for k, v in vars(opt).items()))
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)