capradeepgujaran
commited on
Commit
•
17e6c9d
1
Parent(s):
17991a3
Update app.py
Browse files
app.py
CHANGED
@@ -13,9 +13,19 @@ from tqdm.auto import tqdm
|
|
13 |
from pathlib import Path
|
14 |
from typing import List, Dict, Tuple
|
15 |
import time
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
class VideoProcessor:
|
17 |
def __init__(self):
|
18 |
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
19 |
|
20 |
# Load models with optimizations
|
21 |
self.load_models()
|
@@ -27,25 +37,42 @@ class VideoProcessor:
|
|
27 |
self.batch_size = 4 if torch.cuda.is_available() else 2
|
28 |
|
29 |
def load_models(self):
|
30 |
-
"""Load models with optimizations"""
|
31 |
-
|
32 |
self.clip_model = CLIPModel.from_pretrained(
|
33 |
"openai/clip-vit-base-patch32",
|
34 |
-
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
|
|
|
35 |
).to(self.device)
|
36 |
-
self.clip_processor = CLIPProcessor.from_pretrained(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
-
# Load BLIP2 with
|
39 |
self.blip_model = Blip2ForConditionalGeneration.from_pretrained(
|
40 |
-
|
41 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
42 |
-
device_map="auto" if torch.cuda.is_available() else None
|
|
|
|
|
43 |
).to(self.device)
|
44 |
-
self.blip_processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
|
45 |
|
46 |
# Set models to evaluation mode
|
47 |
self.clip_model.eval()
|
48 |
self.blip_model.eval()
|
|
|
49 |
|
50 |
@torch.no_grad()
|
51 |
def process_frame_batch(self, frames):
|
@@ -55,16 +82,37 @@ class VideoProcessor:
|
|
55 |
pil_frames = [Image.fromarray(cv2.cvtColor(f, cv2.COLOR_BGR2RGB)).resize(self.target_size) for f in frames]
|
56 |
|
57 |
# Get CLIP features
|
58 |
-
clip_inputs = self.clip_processor(
|
|
|
|
|
|
|
|
|
|
|
59 |
if self.device.type == "cuda":
|
60 |
clip_inputs = {k: v.half() if v.dtype == torch.float32 else v for k, v in clip_inputs.items()}
|
61 |
features = self.clip_model.get_image_features(**clip_inputs)
|
62 |
|
63 |
-
# Get BLIP captions
|
64 |
-
blip_inputs = self.blip_processor(
|
|
|
|
|
|
|
|
|
|
|
65 |
if self.device.type == "cuda":
|
66 |
blip_inputs = {k: v.half() if v.dtype == torch.float32 else v for k, v in blip_inputs.items()}
|
67 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
captions = [self.blip_processor.decode(c, skip_special_tokens=True) for c in captions]
|
69 |
|
70 |
return features.cpu().numpy(), captions
|
@@ -75,12 +123,15 @@ class VideoProcessor:
|
|
75 |
def process_video(self, video_path: str, progress=gr.Progress()):
|
76 |
"""Process video with batching and progress updates"""
|
77 |
cap = cv2.VideoCapture(video_path)
|
|
|
|
|
|
|
78 |
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
79 |
fps = cap.get(cv2.CAP_PROP_FPS)
|
80 |
|
81 |
# Calculate frames to process
|
82 |
frames_to_process = min(self.max_frames, total_frames // self.frame_interval)
|
83 |
-
progress(0, desc="Initializing...")
|
84 |
|
85 |
features_list = []
|
86 |
frame_data = []
|
@@ -102,6 +153,9 @@ class VideoProcessor:
|
|
102 |
|
103 |
# Process batch when full
|
104 |
if len(current_batch) == self.batch_size or frame_count == total_frames - 1:
|
|
|
|
|
|
|
105 |
features, captions = self.process_frame_batch(current_batch)
|
106 |
|
107 |
if features is not None and captions is not None:
|
@@ -116,13 +170,9 @@ class VideoProcessor:
|
|
116 |
processed_count += len(current_batch)
|
117 |
current_batch = []
|
118 |
batch_positions = []
|
119 |
-
|
120 |
-
# Update progress
|
121 |
-
progress(processed_count / frames_to_process,
|
122 |
-
desc=f"Processing frames... {processed_count}/{frames_to_process}")
|
123 |
|
124 |
frame_count += 1
|
125 |
-
|
126 |
cap.release()
|
127 |
|
128 |
# Create FAISS index
|
@@ -137,7 +187,7 @@ class VideoProcessor:
|
|
137 |
|
138 |
except Exception as e:
|
139 |
cap.release()
|
140 |
-
|
141 |
|
142 |
class VideoQAInterface:
|
143 |
def __init__(self):
|
@@ -145,13 +195,18 @@ class VideoQAInterface:
|
|
145 |
self.frame_index = None
|
146 |
self.frame_data = None
|
147 |
self.processed = False
|
148 |
-
self.current_video_path = None
|
149 |
self.temp_dir = tempfile.mkdtemp()
|
|
|
150 |
|
151 |
def __del__(self):
|
152 |
"""Cleanup temporary files"""
|
153 |
if hasattr(self, 'temp_dir') and os.path.exists(self.temp_dir):
|
154 |
-
|
|
|
|
|
|
|
|
|
155 |
|
156 |
def process_video(self, video_file, progress=gr.Progress()):
|
157 |
"""Process video with progress tracking"""
|
@@ -163,6 +218,7 @@ class VideoQAInterface:
|
|
163 |
temp_video_path = os.path.join(self.temp_dir, "input_video.mp4")
|
164 |
shutil.copy2(video_file.name, temp_video_path)
|
165 |
self.current_video_path = temp_video_path
|
|
|
166 |
|
167 |
progress(0, desc="Starting video processing...")
|
168 |
self.frame_index, self.frame_data, message = self.processor.process_video(
|
@@ -178,7 +234,7 @@ class VideoQAInterface:
|
|
178 |
|
179 |
except Exception as e:
|
180 |
self.processed = False
|
181 |
-
return f"Error: {str(e)}"
|
182 |
|
183 |
@torch.no_grad()
|
184 |
def answer_question(self, query):
|
@@ -222,7 +278,7 @@ class VideoQAInterface:
|
|
222 |
desc += f"Relevance Score: {result['relevance']:.2f}"
|
223 |
descriptions.append(desc)
|
224 |
finally:
|
225 |
-
cap.release()
|
226 |
|
227 |
if not frames:
|
228 |
return None, "No relevant frames found."
|
@@ -291,4 +347,9 @@ app = VideoQAInterface()
|
|
291 |
interface = app.create_interface()
|
292 |
|
293 |
if __name__ == "__main__":
|
294 |
-
interface.launch(
|
|
|
|
|
|
|
|
|
|
|
|
13 |
from pathlib import Path
|
14 |
from typing import List, Dict, Tuple
|
15 |
import time
|
16 |
+
from huggingface_hub import snapshot_download
|
17 |
+
import warnings
|
18 |
+
warnings.filterwarnings("ignore")
|
19 |
+
|
20 |
+
# Configure model caching and environment
|
21 |
+
os.environ["TRANSFORMERS_CACHE"] = "./model_cache"
|
22 |
+
os.environ["HF_HOME"] = "./model_cache"
|
23 |
+
os.makedirs("./model_cache", exist_ok=True)
|
24 |
+
|
25 |
class VideoProcessor:
|
26 |
def __init__(self):
|
27 |
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
28 |
+
print(f"Using device: {self.device}")
|
29 |
|
30 |
# Load models with optimizations
|
31 |
self.load_models()
|
|
|
37 |
self.batch_size = 4 if torch.cuda.is_available() else 2
|
38 |
|
39 |
def load_models(self):
|
40 |
+
"""Load models with optimizations and proper configurations"""
|
41 |
+
print("Loading CLIP model...")
|
42 |
self.clip_model = CLIPModel.from_pretrained(
|
43 |
"openai/clip-vit-base-patch32",
|
44 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
45 |
+
cache_dir="./model_cache"
|
46 |
).to(self.device)
|
47 |
+
self.clip_processor = CLIPProcessor.from_pretrained(
|
48 |
+
"openai/clip-vit-base-patch32",
|
49 |
+
cache_dir="./model_cache"
|
50 |
+
)
|
51 |
+
|
52 |
+
print("Loading BLIP2 model...")
|
53 |
+
model_name = "Salesforce/blip2-opt-2.7b"
|
54 |
+
|
55 |
+
# Initialize BLIP2 processor with updated configuration
|
56 |
+
self.blip_processor = Blip2Processor.from_pretrained(
|
57 |
+
model_name,
|
58 |
+
cache_dir="./model_cache"
|
59 |
+
)
|
60 |
+
self.blip_processor.config.use_fast_tokenizer = True
|
61 |
+
self.blip_processor.config.processor_class = "Blip2Processor"
|
62 |
|
63 |
+
# Load BLIP2 model with optimizations
|
64 |
self.blip_model = Blip2ForConditionalGeneration.from_pretrained(
|
65 |
+
model_name,
|
66 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
67 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
68 |
+
cache_dir="./model_cache",
|
69 |
+
low_cpu_mem_usage=True
|
70 |
).to(self.device)
|
|
|
71 |
|
72 |
# Set models to evaluation mode
|
73 |
self.clip_model.eval()
|
74 |
self.blip_model.eval()
|
75 |
+
print("Models loaded successfully!")
|
76 |
|
77 |
@torch.no_grad()
|
78 |
def process_frame_batch(self, frames):
|
|
|
82 |
pil_frames = [Image.fromarray(cv2.cvtColor(f, cv2.COLOR_BGR2RGB)).resize(self.target_size) for f in frames]
|
83 |
|
84 |
# Get CLIP features
|
85 |
+
clip_inputs = self.clip_processor(
|
86 |
+
images=pil_frames,
|
87 |
+
return_tensors="pt",
|
88 |
+
padding=True
|
89 |
+
).to(self.device)
|
90 |
+
|
91 |
if self.device.type == "cuda":
|
92 |
clip_inputs = {k: v.half() if v.dtype == torch.float32 else v for k, v in clip_inputs.items()}
|
93 |
features = self.clip_model.get_image_features(**clip_inputs)
|
94 |
|
95 |
+
# Get BLIP captions with updated processing
|
96 |
+
blip_inputs = self.blip_processor(
|
97 |
+
images=pil_frames,
|
98 |
+
return_tensors="pt",
|
99 |
+
padding=True
|
100 |
+
).to(self.device)
|
101 |
+
|
102 |
if self.device.type == "cuda":
|
103 |
blip_inputs = {k: v.half() if v.dtype == torch.float32 else v for k, v in blip_inputs.items()}
|
104 |
+
|
105 |
+
# Generate captions with better parameters
|
106 |
+
captions = self.blip_model.generate(
|
107 |
+
**blip_inputs,
|
108 |
+
max_length=30,
|
109 |
+
min_length=10,
|
110 |
+
num_beams=5,
|
111 |
+
length_penalty=1.0,
|
112 |
+
temperature=0.7,
|
113 |
+
do_sample=False
|
114 |
+
)
|
115 |
+
|
116 |
captions = [self.blip_processor.decode(c, skip_special_tokens=True) for c in captions]
|
117 |
|
118 |
return features.cpu().numpy(), captions
|
|
|
123 |
def process_video(self, video_path: str, progress=gr.Progress()):
|
124 |
"""Process video with batching and progress updates"""
|
125 |
cap = cv2.VideoCapture(video_path)
|
126 |
+
if not cap.isOpened():
|
127 |
+
raise ValueError("Could not open video file")
|
128 |
+
|
129 |
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
130 |
fps = cap.get(cv2.CAP_PROP_FPS)
|
131 |
|
132 |
# Calculate frames to process
|
133 |
frames_to_process = min(self.max_frames, total_frames // self.frame_interval)
|
134 |
+
progress(0, desc="Initializing video processing...")
|
135 |
|
136 |
features_list = []
|
137 |
frame_data = []
|
|
|
153 |
|
154 |
# Process batch when full
|
155 |
if len(current_batch) == self.batch_size or frame_count == total_frames - 1:
|
156 |
+
progress(processed_count / frames_to_process,
|
157 |
+
desc=f"Processing frames... {processed_count}/{frames_to_process}")
|
158 |
+
|
159 |
features, captions = self.process_frame_batch(current_batch)
|
160 |
|
161 |
if features is not None and captions is not None:
|
|
|
170 |
processed_count += len(current_batch)
|
171 |
current_batch = []
|
172 |
batch_positions = []
|
|
|
|
|
|
|
|
|
173 |
|
174 |
frame_count += 1
|
175 |
+
|
176 |
cap.release()
|
177 |
|
178 |
# Create FAISS index
|
|
|
187 |
|
188 |
except Exception as e:
|
189 |
cap.release()
|
190 |
+
raise e
|
191 |
|
192 |
class VideoQAInterface:
|
193 |
def __init__(self):
|
|
|
195 |
self.frame_index = None
|
196 |
self.frame_data = None
|
197 |
self.processed = False
|
198 |
+
self.current_video_path = None
|
199 |
self.temp_dir = tempfile.mkdtemp()
|
200 |
+
print(f"Initialized temp directory: {self.temp_dir}")
|
201 |
|
202 |
def __del__(self):
|
203 |
"""Cleanup temporary files"""
|
204 |
if hasattr(self, 'temp_dir') and os.path.exists(self.temp_dir):
|
205 |
+
try:
|
206 |
+
shutil.rmtree(self.temp_dir)
|
207 |
+
print(f"Cleaned up temp directory: {self.temp_dir}")
|
208 |
+
except Exception as e:
|
209 |
+
print(f"Error cleaning up temp directory: {str(e)}")
|
210 |
|
211 |
def process_video(self, video_file, progress=gr.Progress()):
|
212 |
"""Process video with progress tracking"""
|
|
|
218 |
temp_video_path = os.path.join(self.temp_dir, "input_video.mp4")
|
219 |
shutil.copy2(video_file.name, temp_video_path)
|
220 |
self.current_video_path = temp_video_path
|
221 |
+
print(f"Saved video to: {self.current_video_path}")
|
222 |
|
223 |
progress(0, desc="Starting video processing...")
|
224 |
self.frame_index, self.frame_data, message = self.processor.process_video(
|
|
|
234 |
|
235 |
except Exception as e:
|
236 |
self.processed = False
|
237 |
+
return f"Error processing video: {str(e)}"
|
238 |
|
239 |
@torch.no_grad()
|
240 |
def answer_question(self, query):
|
|
|
278 |
desc += f"Relevance Score: {result['relevance']:.2f}"
|
279 |
descriptions.append(desc)
|
280 |
finally:
|
281 |
+
cap.release()
|
282 |
|
283 |
if not frames:
|
284 |
return None, "No relevant frames found."
|
|
|
347 |
interface = app.create_interface()
|
348 |
|
349 |
if __name__ == "__main__":
|
350 |
+
interface.launch(
|
351 |
+
server_name="0.0.0.0",
|
352 |
+
share=False, # Set to True if you want to create a public link
|
353 |
+
cache_examples=True,
|
354 |
+
max_threads=4
|
355 |
+
)
|