Spaces:
Runtime error
Runtime error
added video analyser modal
Browse files- deploy_serverless_modal_image_processing.py +123 -0
- load_vision_model_locally.py +69 -0
- utils.py +53 -0
deploy_serverless_modal_image_processing.py
ADDED
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import modal
|
2 |
+
from PIL import Image
|
3 |
+
import torch
|
4 |
+
|
5 |
+
# --- Modal App Setup ---
|
6 |
+
# This section defines the environment and models for our serverless functions.
|
7 |
+
# The container image will have all necessary libraries pre-installed.
|
8 |
+
modal_app = modal.App("video-analysis-app")
|
9 |
+
image = modal.Image.debian_slim().pip_install(
|
10 |
+
"torch", "transformers", "Pillow"
|
11 |
+
)
|
12 |
+
|
13 |
+
# Define a class to load models only once when the container starts.
|
14 |
+
# This is a Modal best practice that avoids slow cold starts.
|
15 |
+
# We request a GPU to accelerate the model inference.
|
16 |
+
@modal_app.cls(gpu="any", image=image)
|
17 |
+
class VideoAnalyzer:
|
18 |
+
def __enter__(self):
|
19 |
+
"""
|
20 |
+
This method is called once when the container starts.
|
21 |
+
It loads the models and processors into memory, so they are ready
|
22 |
+
for immediate use by the functions.
|
23 |
+
"""
|
24 |
+
from transformers import BlipProcessor, BlipForConditionalGeneration, DetrImageProcessor, DetrForObjectDetection
|
25 |
+
|
26 |
+
print("Loading Image Captioning model...")
|
27 |
+
# Model for describing the frame (image captioning)
|
28 |
+
self.caption_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
|
29 |
+
self.caption_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large").to("cuda")
|
30 |
+
print("Image Captioning model loaded.")
|
31 |
+
|
32 |
+
print("Loading Object Detection model...")
|
33 |
+
# Model for detecting objects
|
34 |
+
self.detection_processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
|
35 |
+
self.detection_model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50").to("cuda")
|
36 |
+
print("Object Detection model loaded.")
|
37 |
+
|
38 |
+
@modal.method()
|
39 |
+
def describe_frame(self, image_bytes: bytes) -> str:
|
40 |
+
"""
|
41 |
+
Takes image bytes as input and returns a generated text description.
|
42 |
+
"""
|
43 |
+
try:
|
44 |
+
# Open the image from the raw bytes
|
45 |
+
raw_image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
|
46 |
+
|
47 |
+
# Process the image and generate a caption
|
48 |
+
inputs = self.caption_processor(raw_image, return_tensors="pt").to("cuda")
|
49 |
+
out = self.caption_model.generate(**inputs, max_new_tokens=50)
|
50 |
+
|
51 |
+
# Decode the caption and return it
|
52 |
+
caption = self.caption_processor.decode(out[0], skip_special_tokens=True)
|
53 |
+
return caption
|
54 |
+
except Exception as e:
|
55 |
+
print(f"Error describing frame: {e}")
|
56 |
+
return "Could not describe the image."
|
57 |
+
|
58 |
+
@modal.method()
|
59 |
+
def detect_objects(self, image_bytes: bytes, threshold: float = 0.9) -> list[str]:
|
60 |
+
"""
|
61 |
+
Takes image bytes as input and returns a list of detected object labels.
|
62 |
+
Only returns objects with a confidence score above the threshold.
|
63 |
+
"""
|
64 |
+
try:
|
65 |
+
image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
|
66 |
+
|
67 |
+
# Process the image for the object detection model
|
68 |
+
inputs = self.detection_processor(images=image, return_tensors="pt").to("cuda")
|
69 |
+
outputs = self.detection_model(**inputs)
|
70 |
+
|
71 |
+
# Post-process the results to get labels and scores
|
72 |
+
target_sizes = torch.tensor([image.size[::-1]])
|
73 |
+
results = self.detection_processor.post_process_object_detection(
|
74 |
+
outputs, target_sizes=target_sizes, threshold=threshold
|
75 |
+
)[0]
|
76 |
+
|
77 |
+
# Extract the labels for confident detections
|
78 |
+
detected_objects = []
|
79 |
+
for score, label in zip(results["scores"], results["labels"]):
|
80 |
+
object_name = self.detection_model.config.id2label[label.item()]
|
81 |
+
detected_objects.append(object_name)
|
82 |
+
|
83 |
+
# Return a list of unique object names
|
84 |
+
return list(set(detected_objects))
|
85 |
+
except Exception as e:
|
86 |
+
print(f"Error detecting objects: {e}")
|
87 |
+
return []
|
88 |
+
|
89 |
+
# --- Local Runner ---
|
90 |
+
# This part of the script runs on your local machine to call the serverless functions.
|
91 |
+
# It reads a local image file and sends it to the Modal functions for processing.
|
92 |
+
@modal_app.local_entrypoint()
|
93 |
+
def main(image_path: str):
|
94 |
+
"""
|
95 |
+
A local entrypoint to test the Modal functions.
|
96 |
+
|
97 |
+
Usage:
|
98 |
+
modal run this_script_name.py --image-path /path/to/your/image.jpg
|
99 |
+
"""
|
100 |
+
import io
|
101 |
+
|
102 |
+
try:
|
103 |
+
with open(image_path, "rb") as f:
|
104 |
+
img_bytes = f.read()
|
105 |
+
except FileNotFoundError:
|
106 |
+
print(f"Error: Image file not found at '{image_path}'")
|
107 |
+
return
|
108 |
+
|
109 |
+
print("--- Calling Modal Functions ---")
|
110 |
+
|
111 |
+
# Instantiate the class, which will trigger the __enter__ method on the remote container
|
112 |
+
analyzer = VideoAnalyzer()
|
113 |
+
|
114 |
+
# Call the remote functions with the image data
|
115 |
+
description = analyzer.describe_frame.remote(img_bytes)
|
116 |
+
objects = analyzer.detect_objects.remote(img_bytes)
|
117 |
+
|
118 |
+
print(f"\n📸 Analysis for: {image_path}")
|
119 |
+
print("-----------------------------------")
|
120 |
+
print(f"📝 Description: {description}")
|
121 |
+
print(f"📦 Detected Objects: {objects}")
|
122 |
+
print("-----------------------------------")
|
123 |
+
|
load_vision_model_locally.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import cv2
|
2 |
+
import os
|
3 |
+
import io
|
4 |
+
from PIL import Image
|
5 |
+
import torch
|
6 |
+
from transformers import BlipProcessor, BlipForConditionalGeneration, DetrImageProcessor, DetrForObjectDetection
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
class VideoAnalyzer:
|
11 |
+
def __init__(self):
|
12 |
+
"""Initialize the models."""
|
13 |
+
print("Loading Image Captioning model...")
|
14 |
+
self.caption_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
|
15 |
+
self.caption_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large")
|
16 |
+
|
17 |
+
print("Loading Object Detection model...")
|
18 |
+
self.detection_processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
|
19 |
+
self.detection_model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
|
20 |
+
|
21 |
+
def describe_frame(self, image_path: str) -> str:
|
22 |
+
"""Generate a text description of the frame."""
|
23 |
+
try:
|
24 |
+
raw_image = Image.open(image_path).convert("RGB")
|
25 |
+
inputs = self.caption_processor(raw_image, return_tensors="pt")
|
26 |
+
out = self.caption_model.generate(**inputs, max_new_tokens=50)
|
27 |
+
caption = self.caption_processor.decode(out[0], skip_special_tokens=True)
|
28 |
+
return caption
|
29 |
+
except Exception as e:
|
30 |
+
print(f"Error describing frame: {e}")
|
31 |
+
return "Could not describe the image."
|
32 |
+
|
33 |
+
def detect_objects(self, image_path: str, threshold: float = 0.9) -> list[str]:
|
34 |
+
"""Detect objects in the frame."""
|
35 |
+
try:
|
36 |
+
image = Image.open(image_path).convert("RGB")
|
37 |
+
inputs = self.detection_processor(images=image, return_tensors="pt")
|
38 |
+
outputs = self.detection_model(**inputs)
|
39 |
+
target_sizes = torch.tensor([image.size[::-1]])
|
40 |
+
results = self.detection_processor.post_process_object_detection(
|
41 |
+
outputs, target_sizes=target_sizes, threshold=threshold
|
42 |
+
)[0]
|
43 |
+
|
44 |
+
detected_objects = []
|
45 |
+
for score, label in zip(results["scores"], results["labels"]):
|
46 |
+
object_name = self.detection_model.config.id2label[label.item()]
|
47 |
+
detected_objects.append(object_name)
|
48 |
+
|
49 |
+
return list(set(detected_objects))
|
50 |
+
except Exception as e:
|
51 |
+
print(f"Error detecting objects: {e}")
|
52 |
+
return []
|
53 |
+
|
54 |
+
# Initialize the VideoAnalyzer
|
55 |
+
analyzer = VideoAnalyzer()
|
56 |
+
|
57 |
+
def get_frame_infos(filename: str) -> dict:
|
58 |
+
"""Extract information from a frame."""
|
59 |
+
if not os.path.exists(filename):
|
60 |
+
return {"error": "File not found"}
|
61 |
+
|
62 |
+
description = analyzer.describe_frame(filename)
|
63 |
+
objects = analyzer.detect_objects(filename)
|
64 |
+
|
65 |
+
return {
|
66 |
+
"filename": filename,
|
67 |
+
"description": description,
|
68 |
+
"objects": objects
|
69 |
+
}
|
utils.py
CHANGED
@@ -124,3 +124,56 @@ def extract_keyframes(video_path, diff_threshold=0.4):
|
|
124 |
cap.release()
|
125 |
print(f"Extracted {saved_id} key frames.")
|
126 |
return "success"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
cap.release()
|
125 |
print(f"Extracted {saved_id} key frames.")
|
126 |
return "success"
|
127 |
+
|
128 |
+
|
129 |
+
def extract_nfps_frames(video_path, nfps=5,diff_threshold=0.4):
|
130 |
+
"""Extract 1 frame per second from a video.
|
131 |
+
Args:
|
132 |
+
video_path (str): Path to the input video file.
|
133 |
+
"""
|
134 |
+
cap = cv2.VideoCapture(video_path)
|
135 |
+
if not cap.isOpened():
|
136 |
+
print("Failed to read video.")
|
137 |
+
return
|
138 |
+
|
139 |
+
output_path = '/tmp/video/frames'
|
140 |
+
os.makedirs(output_path, exist_ok=True)
|
141 |
+
|
142 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
143 |
+
frame_interval = int(fps) * nfps # Capture one frame every n second
|
144 |
+
|
145 |
+
frame_id = 0
|
146 |
+
saved_id = 0
|
147 |
+
success, prev_frame = cap.read()
|
148 |
+
|
149 |
+
while True:
|
150 |
+
success, frame = cap.read()
|
151 |
+
if not success:
|
152 |
+
break
|
153 |
+
|
154 |
+
if frame_id % frame_interval == 0 and is_significantly_different(prev_frame, frame, threshold=diff_threshold):
|
155 |
+
filename = os.path.join(output_path, f"frame_{saved_id:04d}.jpg")
|
156 |
+
cv2.imwrite(filename, frame)
|
157 |
+
prev_frame = frame
|
158 |
+
saved_id += 1
|
159 |
+
|
160 |
+
# append to a list that will constitute RAG Docuement
|
161 |
+
frame_data=get_frame_infos(filename)
|
162 |
+
all_frames_data.append(frame_data)
|
163 |
+
frame_id += 1
|
164 |
+
|
165 |
+
cap.release()
|
166 |
+
print(f"Extracted {saved_id} frames (1 per second).")
|
167 |
+
return all_frames_data
|
168 |
+
|
169 |
+
def get_frame_infos(filename:str) -> dict:
|
170 |
+
from load_vision_model_locally import VideoAnalyser
|
171 |
+
analyser = VideoAnalyser()
|
172 |
+
|
173 |
+
description = analyser.describe_frame(filename)
|
174 |
+
detection = analyser.detect_objects(filename)
|
175 |
+
|
176 |
+
print("description",type(description),description)
|
177 |
+
print("detection",type(detection),detection)
|
178 |
+
|
179 |
+
return (descrition, detection)
|