File size: 11,747 Bytes
ede298f 0a90baa ede298f 0a90baa ede298f 46d3efb ede298f 0a90baa ede298f 0a90baa ede298f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 |
import argparse
import os
import yaml
from pathlib import Path
from datetime import datetime
import logging
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim.lr_scheduler import StepLR
from torch.utils.data import DataLoader
from torchvision import transforms
import wandb
from sklearn.metrics import confusion_matrix
import laspy
from datasets.semantic_dataset import SemanticSegmentationDataset
from models.upernet_wrapper import load_model
from train.train import train_epoch
from train.eval import eval_epoch
from inference.inference import inference
from inference.export_logits import export_logits_images
from projection.lidar_projection import project_classes_to_lidar, project_softmax_to_lidar
from utils.seed import set_seed
from utils.logging_utils import init_logging
from utils.metrics import compute_metrics
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=Path, default="config.yaml", help="Path to YAML configuration.")
parser.add_argument("--mode", choices=["train", "val", "val3d", "test", "test3d", "export_probs", "project_probs_3d"], default="train")
parser.add_argument("--weights_path", default="best_model.pth", help="Path to model weights.")
args = parser.parse_args()
cfg = yaml.safe_load(args.config.read_text())
class_map = {tuple(item["keys"]): item["value"] for item in cfg["data"].get("class_map", [])}
now = datetime.now()
out_dir = Path(cfg["training"]["output_dir"] + now.strftime("_%m_%d_%H_%M_%S"))
out_dir.mkdir(parents=True, exist_ok=True)
init_logging(out_dir / "run.log")
set_seed(cfg["training"]["seed"])
batch_size_projection = int(cfg["val"]["batch_size_proj"])
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(cfg["data"]["mean"], cfg["data"]["std"])
])
ds_args = dict(
root_dir=Path(cfg["data"]["root_dir"]),
split_json=Path(cfg["data"]["split_file"]),
resize_size=tuple(cfg["data"]["resize_size"]),
crop_size=tuple(cfg["data"]["crop_size"]),
class_map=class_map,
transform=transform
)
mode = args.mode
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = load_model(cfg["model"]["pretrained_model"], cfg["model"]["num_classes"], device)
criterion = nn.CrossEntropyLoss(ignore_index=cfg["model"]["ignore_index"])
optimizer = optim.AdamW(model.parameters(), lr=cfg["training"]["lr"])
scheduler = StepLR(optimizer, step_size=cfg["training"]["sched_step"], gamma=cfg["training"]["sched_gamma"])
scaler = torch.cuda.amp.GradScaler()
if mode in ["train", "val"]:
wandb.init(project=cfg["wandb"]["project"], entity=cfg["wandb"]["entity"], config=cfg)
wandb.watch(model, log="all")
train_ds = SemanticSegmentationDataset(split="train", mode=mode, **ds_args)
val_ds = SemanticSegmentationDataset(split="val", mode=mode, **ds_args)
train_loader = DataLoader(train_ds, batch_size=cfg["training"]["batch_size"],
shuffle=True, num_workers=cfg["training"]["num_workers"],
pin_memory=True, persistent_workers=True)
val_loader = DataLoader(val_ds, batch_size=cfg["val"]["batch_size"],
shuffle=False, num_workers=cfg["val"]["num_workers"],
pin_memory=True, persistent_workers=True)
if mode in ["test", "test3d"]:
test_ds = SemanticSegmentationDataset(split="test", mode=mode, **ds_args)
test_loader = DataLoader(test_ds, batch_size=cfg["val"]["batch_size"],
shuffle=False, num_workers=cfg["val"]["num_workers"],
pin_memory=True, persistent_workers=True)
if mode == "val3d":
wandb.init(project=cfg["wandb"]["project"], entity=cfg["wandb"]["entity"], config=cfg)
wandb.watch(model, log="all")
val_ds = SemanticSegmentationDataset(split="val", mode=mode, **ds_args)
val_loader = DataLoader(val_ds, batch_size=cfg["val"]["batch_size"],
shuffle=False, num_workers=cfg["val"]["num_workers"],
pin_memory=True, persistent_workers=True)
if mode == "train":
best_miou = 0.0
eval_every = cfg["training"].get("eval_every", 1)
for epoch in range(1, cfg["training"]["epochs"] + 1):
logging.info(f"Epoch {epoch}/{cfg['training']['epochs']}")
train_loss = train_epoch(model, train_loader, criterion, optimizer, device, scaler)
scheduler.step()
if epoch % eval_every == 0 or epoch == cfg["training"]["epochs"]:
ious, miou, mf1 = eval_epoch(model, val_loader, criterion, device,
ds_args['crop_size'], ds_args['crop_size'],
cfg["model"]["num_classes"])
wandb.log({
"epoch": epoch, "train_loss": train_loss,
"val_mIoU": miou, "val_mF1": mf1,
"lr": optimizer.param_groups[0]["lr"],
**{f"val_iou_class_{i}": v for i, v in enumerate(ious)}
})
logging.info(f"Train Loss: {train_loss:.4f} | Val mIoU: {miou:.4f} | Val mF1: {mf1:.4f}")
if miou > best_miou:
best_miou = miou
torch.save(model.state_dict(), out_dir / "best_model.pth")
torch.save(model.state_dict(), out_dir / f"model_epoch{epoch:02d}.pth")
elif mode == "val":
model.load_state_dict(torch.load(args.weights_path))
ious, miou, mf1 = eval_epoch(model, val_loader, criterion, device,
ds_args['crop_size'], ds_args['crop_size'],
cfg["model"]["num_classes"])
logging.info(f"Validation | mIoU: {miou:.4f} | mF1: {mf1:.4f} | IoUs: {ious}")
wandb.log({
"val_mIoU": miou, "val_mF1": mf1,
**{f"val_iou_class_{i}": v for i, v in enumerate(ious)}
})
elif mode in ["test", "test3d"]:
model.load_state_dict(torch.load(args.weights_path))
inference(model, test_loader, device, ds_args["crop_size"], ds_args["crop_size"],
out_dir / "predictions")
logging.info(f"Inference Image complete. Predictions saved to {out_dir/'predictions'}")
if mode == "test3d":
for zone in sorted(os.listdir(out_dir / "predictions")):
if not (out_dir / "predictions" / zone).is_dir():
continue
project_classes_to_lidar(zone_name=zone,
dataset_root=cfg["data"]["root_dir"],
outputs_root=str(out_dir / "predictions"),
nb_classes=cfg["model"]["num_classes"],
output_path=str(out_dir / "predictions" / zone / f"{zone}_with_classif.las"),
batch_size=batch_size_projection)
elif mode == "val3d":
model.load_state_dict(torch.load(args.weights_path))
inference(model, val_loader, device, ds_args["crop_size"], ds_args["crop_size"],
out_dir / "predictions")
logging.info(f"Inference Image complete. Predictions saved to {out_dir/'predictions'}")
conf_mat = np.zeros((cfg["model"]["num_classes"], cfg["model"]["num_classes"]), dtype=int)
for zone in sorted(os.listdir(out_dir / "predictions")):
output_las_path = out_dir / "predictions" / zone / f"{zone}_with_classif.las"
if not output_las_path.exists():
project_classes_to_lidar(zone_name=zone,
dataset_root=cfg["data"]["root_dir"],
outputs_root=str(out_dir / "predictions"),
nb_classes=cfg["model"]["num_classes"],
output_path=str(output_las_path),
batch_size=batch_size_projection)
las_pred = laspy.read(output_las_path)
las_gt = laspy.read(Path(cfg["data"]["root_dir"]) / zone / "lidar" / f"{zone}.las")
lut = np.full(256, fill_value=255, dtype=np.uint8)
for keys, v in class_map.items():
lut[list(keys)] = v
pred = np.array(las_pred.classif, dtype=np.int32)
gt = lut[np.array(las_gt.ground_truth, dtype=np.int32)]
valid = gt != cfg["model"]["ignore_index"]
conf_mat += confusion_matrix(gt[valid], pred[valid], labels=list(range(cfg["model"]["num_classes"])))
ious, miou, _, mf1 = compute_metrics(conf_mat)
logging.info(f"Validation 3D | mIoU: {miou:.4f} | mF1: {mf1:.4f} | IoUs: {ious}")
wandb.log({
"val_mIoU_3D": miou, "val_mF1_3D": mf1,
**{f"val_iou_class_{i}_3D": v for i, v in enumerate(ious)}
})
elif mode == "export_probs":
model.load_state_dict(torch.load(args.weights_path))
model.to(device)
for split in ["train", "val", "test"]:
ds = SemanticSegmentationDataset(split=split, mode="test", **ds_args)
loader = DataLoader(ds, batch_size=cfg["val"]["batch_size"],
shuffle=False, num_workers=cfg["val"]["num_workers"],
pin_memory=True, persistent_workers=True)
out_logits_dir = out_dir / f"logits_{split}"
export_logits_images(model, loader, device, ds_args["crop_size"],
ds_args["crop_size"], out_logits_dir)
logging.info(f"Logits exported for split={split} in {out_logits_dir}")
elif mode == "project_probs_3d":
model.load_state_dict(torch.load(args.weights_path))
model.to(device)
for split in ["train", "val", "test"]:
ds = SemanticSegmentationDataset(
split=split,
mode="test", # to obtain rel_paths
**ds_args
)
loader = DataLoader(
ds,
batch_size=cfg["val"]["batch_size"],
shuffle=False,
num_workers=cfg["val"]["num_workers"],
pin_memory=True,
persistent_workers=True
)
out_logits_dir = out_dir / f"logits_{split}"
export_logits_images(
model=model,
loader=loader,
device=device,
crop_size=ds_args["crop_size"],
stride=ds_args["crop_size"],
output_dir=out_logits_dir
)
# Projection in las
dataset_root = Path(cfg["data"]["root_dir"])
nb_classes = cfg["model"]["num_classes"]
for zone in sorted(os.listdir(out_logits_dir)):
zone_path = out_logits_dir / zone
if not zone_path.is_dir():
continue
output_las_path = out_logits_dir / zone / f"{zone}_with_softmax.las"
project_softmax_to_lidar(
zone_name=zone,
dataset_root=dataset_root,
logits_root=out_logits_dir,
nb_classes=nb_classes,
output_path=str(output_las_path),
batch_size=batch_size_projection
)
if __name__ == "__main__":
main()
|