File size: 13,808 Bytes
83e35a7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 |
# Cell 1
import torch
from torchvision import transforms
from PIL import Image
import numpy as np
from backend.keyframes.model import DSN
import torch.nn as nn
import cv2
import time
import os
import srt
from backend.keyframes.extract_frames import extract_frames
from backend.utils import copy_and_rename_file, get_black_bar_coordinates, crop_image
import signal
import threading # Added to check main thread
# Cell 2
# Global model cache to avoid reloading
_googlenet_model = None
_preprocess_pipeline = None
def _get_features(frames, gpu=True, batch_size=1):
global _googlenet_model, _preprocess_pipeline
# Load pre-trained GoogLeNet model only once
if _googlenet_model is None:
print("π Loading GoogLeNet model (this happens only once)...")
_googlenet_model = torch.hub.load('pytorch/vision:v0.10.0', 'googlenet', weights='GoogLeNet_Weights.DEFAULT')
# Remove the classification layer (last layer) to obtain features
_googlenet_model = torch.nn.Sequential(*(list(_googlenet_model.children())[:-1]))
_googlenet_model.eval()
# Initialize preprocessing pipeline
_preprocess_pipeline = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
# Move to GPU if available
if gpu:
_googlenet_model.to('cuda')
print("β
GoogLeNet model loaded successfully")
# Initialize a list to store the features
features = []
# Iterate through frames
for frame_path in frames:
# Load and preprocess the frame
input_image = Image.open(frame_path)
input_tensor = _preprocess_pipeline(input_image)
input_batch = input_tensor.unsqueeze(0) # Add batch dimension
# Move the input to GPU if available
if gpu:
input_batch = input_batch.to('cuda')
# Perform feature extraction
with torch.no_grad():
output = _googlenet_model(input_batch)
# Append the features to the list
features.append(output.squeeze().cpu().numpy())
# Convert the list of features to a NumPy array
features = np.array(features)
return features.astype(np.float32)
# Global DSN model cache
_dsn_models = {}
def _get_probs(features, gpu=True, mode=0):
global _dsn_models
# Create cache key
cache_key = f"dsn_model_{mode}_{gpu}"
# Load model only if not already cached
if cache_key not in _dsn_models:
print(f"π Loading DSN model {mode} (this happens only once)...")
if mode == 1:
model_path = "backend/keyframes/pretrained_model/model_1.pth.tar"
else:
model_path = "backend/keyframes/pretrained_model/model_0.pth.tar"
model = DSN(in_dim=1024, hid_dim=256, num_layers=1, cell="lstm")
if gpu:
checkpoint = torch.load(model_path)
else:
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint)
if gpu:
model = nn.DataParallel(model).cuda()
model.eval()
_dsn_models[cache_key] = model
print(f"β
DSN model {mode} loaded successfully")
model = _dsn_models[cache_key]
seq = torch.from_numpy(features).unsqueeze(0)
if gpu: seq = seq.cuda()
probs = model(seq)
probs = probs.data.cpu().squeeze().numpy()
return probs
def generate_keyframes(video):
data=""
with open("test1.srt") as f:
data = f.read()
subs = srt.parse(data)
torch.cuda.empty_cache()
# Add timeout protection
def timeout_handler(signum, frame):
raise TimeoutError("Keyframe generation timed out")
# Set timeout to 10 minutes only if running in the main thread (signals are not allowed in worker threads)
if threading.current_thread() is threading.main_thread():
signal.signal(signal.SIGALRM, timeout_handler)
signal.alarm(600) # 10 minutes timeout
# Create final directory if it doesn't exist
final_dir = os.path.join("frames", "final")
if not os.path.exists(final_dir):
os.makedirs(final_dir)
print(f"Created directory: {final_dir}")
frame_counter = 1
total_subs = len(list(subs))
subs = list(subs) # Convert to list to avoid exhaustion
print(f"π― Processing {total_subs} subtitle segments...")
try:
# Enhanced story-aware keyframe extraction
for i, sub in enumerate(subs, 1):
print(f"π Processing segment {i}/{total_subs}: {sub.content[:30]}...")
frames = []
if not os.path.exists(f"frames/sub{sub.index}"):
os.makedirs(f"frames/sub{sub.index}")
# Extract more frames per segment for better story selection
frames = extract_frames(video, os.path.join("frames", f"sub{sub.index}"),
sub.start.total_seconds(), sub.end.total_seconds(), 10) # Increased from 3 to 10
if len(frames) > 0:
# Get AI highlight scores
features = _get_features(frames, gpu=False)
highlight_scores = _get_probs(features, gpu=False)
# Enhanced story-aware selection
story_frames = _select_story_relevant_frames(frames, highlight_scores, sub)
# Save the best story frames
for j, frame_idx in enumerate(story_frames):
if frame_counter <= 16: # Limit to 16 frames total
try:
copy_and_rename_file(frames[frame_idx], final_dir, f"frame{frame_counter:03}.png")
print(f"π Story frame {frame_counter}: {sub.content} (score: {highlight_scores[frame_idx]:.3f})")
frame_counter += 1
except:
pass
else:
# Fallback if no frames extracted
print(f"β οΈ No frames extracted for subtitle {sub.index}")
# If no frames were successfully generated, run fallback extraction on full video
if frame_counter == 1:
print("π¨ No story-relevant frames generated β falling back to uniform extractionβ¦")
try:
# Extract 16 evenly spaced frames across the entire video duration
video_cap = cv2.VideoCapture(video)
total_frames = int(video_cap.get(cv2.CAP_PROP_FRAME_COUNT))
step = max(total_frames // 16, 1)
extracted = 0
frame_idx = 0
while extracted < 16 and video_cap.isOpened():
video_cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
ret, frame = video_cap.read()
if not ret:
break
out_path = os.path.join(final_dir, f"frame{frame_counter:03}.png")
cv2.imwrite(out_path, frame)
frame_counter += 1
extracted += 1
frame_idx += step
video_cap.release()
print(f"β
Fallback extracted {extracted} uniform frames")
except Exception as e:
print(f"Fallback extraction failed: {e}")
print(f"β
Generated {frame_counter-1} story-relevant frames")
except TimeoutError:
print("β° Keyframe generation timed out, using fallback method...")
# Fallback: use first few subtitle segments
for i, sub in enumerate(subs[:4], 1): # Use only first 4 segments
if frame_counter <= 16:
try:
# Simple frame extraction without AI
frames = extract_frames(video, os.path.join("frames", f"sub{sub.index}"),
sub.start.total_seconds(), sub.end.total_seconds(), 1)
if frames:
copy_and_rename_file(frames[0], final_dir, f"frame{frame_counter:03}.png")
print(f"π Fallback frame {frame_counter}: {sub.content}")
frame_counter += 1
except:
pass
print(f"β
Generated {frame_counter-1} fallback frames")
finally:
# Cancel timeout
signal.alarm(0)
def _select_story_relevant_frames(frames, highlight_scores, subtitle):
"""Enhanced story-aware frame selection"""
try:
highlight_scores = list(highlight_scores)
# 1. Get top AI-scored frames
sorted_indices = [i[0] for i in sorted(enumerate(highlight_scores), key=lambda x: x[1], reverse=True)]
# 2. Analyze frames for story relevance
story_scores = []
for i, frame_path in enumerate(frames):
story_score = _analyze_story_relevance(frame_path, highlight_scores[i], subtitle)
story_scores.append(story_score)
# 3. Combine AI scores with story relevance
combined_scores = []
for i in range(len(frames)):
combined_score = (highlight_scores[i] * 0.6) + (story_scores[i] * 0.4) # 60% AI, 40% story
combined_scores.append(combined_score)
# 4. Select top frames based on combined scores
sorted_combined = [i[0] for i in sorted(enumerate(combined_scores), key=lambda x: x[1], reverse=True)]
# Return top 2-3 frames per segment for better story coverage
num_frames_to_select = min(3, len(frames))
return sorted_combined[:num_frames_to_select]
except Exception as e:
print(f"Story selection failed: {e}")
# Fallback to original method
try:
highlight_scores = list(highlight_scores)
sorted_indices = [i[0] for i in sorted(enumerate(highlight_scores), key=lambda x: x[1], reverse=True)]
return [sorted_indices[0]] if sorted_indices else [0]
except:
return [0] # Ultimate fallback
def _analyze_story_relevance(frame_path, ai_score, subtitle):
"""Analyze frame for story relevance"""
try:
img = cv2.imread(frame_path)
if img is None:
return ai_score
# 1. Face detection (dialogue scenes are important)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
faces = face_cascade.detectMultiScale(gray, 1.1, 4)
face_score = len(faces) * 0.2 # Bonus for faces
# 2. Motion/action detection
motion_score = _detect_motion(img) * 0.15
# 3. Scene complexity (more complex scenes might be more important)
complexity_score = _analyze_scene_complexity(img) * 0.1
# 4. Subtitle content analysis
content_score = _analyze_subtitle_relevance(subtitle.content) * 0.15
# Combine scores
story_score = ai_score + face_score + motion_score + complexity_score + content_score
return min(story_score, 1.0) # Cap at 1.0
except Exception as e:
return ai_score # Fallback to AI score
def _detect_motion(img):
"""Detect motion/action in frame"""
try:
# Simple edge density as motion indicator
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
edges = cv2.Canny(gray, 50, 150)
edge_density = np.sum(edges > 0) / (edges.shape[0] * edges.shape[1])
return min(edge_density * 10, 1.0) # Normalize to 0-1
except:
return 0.0
def _analyze_scene_complexity(img):
"""Analyze scene complexity"""
try:
# Use color variance as complexity indicator
lab = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
l_channel = lab[:,:,0]
complexity = np.std(l_channel) / 255.0
return min(complexity * 2, 1.0) # Normalize to 0-1
except:
return 0.0
def _analyze_subtitle_relevance(subtitle_text):
"""Analyze subtitle content for story relevance"""
# Keywords that indicate important story moments
important_keywords = [
'hello', 'goodbye', 'thank', 'please', 'sorry', 'yes', 'no',
'love', 'hate', 'help', 'danger', 'important', 'secret',
'action', 'fight', 'run', 'stop', 'go', 'come', 'leave'
]
text_lower = subtitle_text.lower()
relevance_score = 0.0
for keyword in important_keywords:
if keyword in text_lower:
relevance_score += 0.1
return min(relevance_score, 1.0) # Cap at 1.0
def black_bar_crop():
ref_img_path = "frames/final/frame001.png"
# Check if reference image exists
if not os.path.exists(ref_img_path):
print(f"β Reference image not found: {ref_img_path}")
return 0, 0, 0, 0
x, y, w, h = get_black_bar_coordinates(ref_img_path)
# Loop through each keyframe
folder_dir = "frames/final"
if not os.path.exists(folder_dir):
print(f"β Frames directory not found: {folder_dir}")
return x, y, w, h
for image in os.listdir(folder_dir):
img_path = os.path.join("frames",'final',image)
if os.path.exists(img_path):
image_data = cv2.imread(img_path)
if image_data is not None:
# Crop the image
crop = image_data[y:y+h, x:x+w]
# Save the cropped image
cv2.imwrite(img_path, crop)
return x, y, w, h |