Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -14,21 +14,20 @@ import faiss
|
|
14 |
logging.basicConfig(level=logging.INFO)
|
15 |
logger = logging.getLogger(__name__)
|
16 |
|
17 |
-
# Check
|
18 |
logger.info(f"CUDA available: {torch.cuda.is_available()}")
|
19 |
if torch.cuda.is_available():
|
20 |
logger.info(f"GPU: {torch.cuda.get_device_name(0)}")
|
21 |
-
logger.info(f"Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9} GB")
|
22 |
|
23 |
-
#
|
24 |
bnb_config = BitsAndBytesConfig(
|
25 |
load_in_4bit=True,
|
26 |
bnb_4bit_quant_type="nf4",
|
27 |
bnb_4bit_compute_dtype=torch.float16,
|
28 |
)
|
29 |
|
|
|
30 |
try:
|
31 |
-
# Load Qwen-2.5-Omni-3B with memory optimizations
|
32 |
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Omni-3B", trust_remote_code=True)
|
33 |
model = AutoModelForCausalLM.from_pretrained(
|
34 |
"Qwen/Qwen2.5-Omni-3B",
|
@@ -36,211 +35,160 @@ try:
|
|
36 |
quantization_config=bnb_config,
|
37 |
trust_remote_code=True
|
38 |
).eval()
|
39 |
-
logger.info("
|
40 |
except Exception as e:
|
41 |
-
logger.error(f"
|
42 |
-
model = None
|
43 |
-
tokenizer = None
|
44 |
|
45 |
-
#
|
46 |
try:
|
47 |
embed_model = SentenceTransformer('paraphrase-MiniLM-L3-v2')
|
48 |
-
logger.info("Embedding model loaded
|
49 |
except Exception as e:
|
50 |
-
logger.error(f"
|
51 |
embed_model = None
|
52 |
|
53 |
-
# Global state
|
54 |
chunks = []
|
55 |
index = None
|
56 |
|
57 |
-
# PDF
|
58 |
def extract_chunks_from_pdf(pdf_path, chunk_size=1000, overlap=200):
|
59 |
try:
|
60 |
doc = fitz.open(pdf_path)
|
61 |
-
text = ""
|
62 |
-
for page in doc:
|
63 |
-
text += page.get_text()
|
64 |
return [text[i:i + chunk_size] for i in range(0, len(text), chunk_size - overlap)]
|
65 |
except Exception as e:
|
66 |
-
logger.error(f"PDF
|
67 |
-
return ["Error extracting
|
68 |
|
|
|
69 |
def build_faiss_index(chunks):
|
70 |
try:
|
71 |
-
if not embed_model:
|
72 |
-
return None
|
73 |
embeddings = embed_model.encode(chunks, convert_to_numpy=True)
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
return idx
|
78 |
except Exception as e:
|
79 |
logger.error(f"FAISS index error: {e}")
|
80 |
return None
|
81 |
|
|
|
82 |
def rag_query(query, chunks, index, top_k=3):
|
83 |
-
if not index or not embed_model:
|
84 |
-
return "Embedding model not available"
|
85 |
try:
|
86 |
q_emb = embed_model.encode([query], convert_to_numpy=True)
|
87 |
D, I = index.search(q_emb, top_k)
|
88 |
return "\n\n".join([chunks[i] for i in I[0]])
|
89 |
except Exception as e:
|
90 |
logger.error(f"RAG query error: {e}")
|
91 |
-
return "Error retrieving context"
|
92 |
|
93 |
-
#
|
94 |
-
def chat_with_qwen(text
|
95 |
if not model or not tokenizer:
|
96 |
-
return "Model
|
97 |
-
|
98 |
try:
|
99 |
-
|
100 |
-
messages = []
|
101 |
-
|
102 |
if image:
|
103 |
-
|
104 |
-
|
105 |
-
{"image": image},
|
106 |
-
{"text": text if text else "Please describe this image."}
|
107 |
-
]})
|
108 |
-
else:
|
109 |
-
# Text-only query
|
110 |
-
messages.append({"role": "user", "content": text})
|
111 |
-
|
112 |
-
# Generate response
|
113 |
-
response = model.chat(tokenizer, messages)
|
114 |
return response
|
115 |
except Exception as e:
|
116 |
logger.error(f"Chat error: {e}")
|
117 |
-
return f"
|
118 |
|
119 |
-
#
|
120 |
def extract_video_frames(video_path, max_frames=2):
|
121 |
try:
|
122 |
cap = cv2.VideoCapture(video_path)
|
123 |
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
|
|
124 |
frames = []
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
|
131 |
-
success, frame = cap.read()
|
132 |
-
if success:
|
133 |
-
frames.append(frame)
|
134 |
cap.release()
|
135 |
return frames
|
136 |
except Exception as e:
|
137 |
-
logger.error(f"
|
138 |
return []
|
139 |
|
140 |
-
#
|
141 |
def multimodal_chat(message, history, image=None, video=None, pdf=None):
|
142 |
global chunks, index
|
143 |
-
|
144 |
if not model:
|
145 |
-
return "Model not
|
146 |
|
147 |
try:
|
148 |
-
# PDF
|
149 |
-
if pdf:
|
150 |
-
|
|
|
|
|
|
|
151 |
index = build_faiss_index(chunks)
|
152 |
if index:
|
153 |
context = rag_query(message, chunks, index)
|
154 |
-
|
155 |
-
|
156 |
else:
|
157 |
-
|
158 |
-
return response
|
159 |
|
160 |
-
# Image
|
161 |
-
if image:
|
162 |
-
|
163 |
-
return response
|
164 |
|
165 |
-
# Video
|
166 |
-
if video:
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
shutil.copy(video, video_path)
|
171 |
frames = extract_video_frames(video_path)
|
|
|
|
|
172 |
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
frame_descriptions = []
|
177 |
-
for i, frame in enumerate(frames):
|
178 |
-
temp_img_path = os.path.join(temp_dir, f"frame_{i}.jpg")
|
179 |
-
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
180 |
-
cv2.imwrite(temp_img_path, frame_rgb)
|
181 |
-
|
182 |
-
# Get description for this frame
|
183 |
-
frame_query = "Describe this video frame in detail."
|
184 |
-
frame_description = chat_with_qwen(frame_query, temp_img_path)
|
185 |
-
frame_descriptions.append(f"Frame {i+1}: {frame_description}")
|
186 |
-
|
187 |
-
# Combine frame descriptions and answer the user's question
|
188 |
-
combined_context = "\n\n".join(frame_descriptions)
|
189 |
-
final_prompt = f"I analyzed some video frames and here's what I found:\n\n{combined_context}\n\nBased on these video frames, {message if message else 'please describe what\'s happening in this video.'}"
|
190 |
-
response = chat_with_qwen(final_prompt)
|
191 |
-
return response
|
192 |
-
else:
|
193 |
-
return "Could not extract video frames"
|
194 |
-
finally:
|
195 |
-
# Cleanup temp files
|
196 |
-
shutil.rmtree(temp_dir, ignore_errors=True)
|
197 |
|
198 |
# Text only
|
199 |
if message:
|
200 |
return chat_with_qwen(message)
|
201 |
|
202 |
-
return "Please
|
203 |
except Exception as e:
|
204 |
-
logger.error(f"
|
205 |
-
return f"Error
|
206 |
|
207 |
-
#
|
208 |
with gr.Blocks(css="""
|
209 |
-
body {
|
210 |
-
|
211 |
-
}
|
212 |
-
.gradio-container {
|
213 |
-
font-family: 'Segoe UI', sans-serif;
|
214 |
-
}
|
215 |
h1 {
|
216 |
-
background: linear-gradient(to right, #667eea, #764ba2);
|
217 |
-
color: white !important;
|
218 |
-
padding: 1rem;
|
219 |
-
border-radius: 12px;
|
220 |
-
margin-bottom: 0.5rem;
|
221 |
-
}
|
222 |
-
p {
|
223 |
-
font-size: 1rem;
|
224 |
-
color: white;
|
225 |
}
|
226 |
.gr-box {
|
227 |
-
background-color: white;
|
228 |
-
|
229 |
-
box-shadow: 0 0 10px rgba(0,0,0,0.05);
|
230 |
-
padding: 16px;
|
231 |
}
|
232 |
-
footer {display: none !important;}
|
233 |
""") as demo:
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
|
|
238 |
|
239 |
chatbot = gr.Chatbot(show_label=False, height=450)
|
240 |
state = gr.State([])
|
241 |
|
242 |
with gr.Row():
|
243 |
-
txt = gr.Textbox(show_label=False, placeholder="Type
|
244 |
send_btn = gr.Button("🚀 Send", scale=1)
|
245 |
|
246 |
with gr.Row():
|
@@ -250,14 +198,13 @@ footer {display: none !important;}
|
|
250 |
|
251 |
def user_send(message, history, image, video, pdf):
|
252 |
if not message and not image and not video and not pdf:
|
253 |
-
return "", history
|
254 |
response = multimodal_chat(message, history, image, video, pdf)
|
255 |
history.append((message, response))
|
256 |
-
return "", history
|
257 |
|
258 |
-
send_btn.click(user_send, [txt, state, image_input, video_input, pdf_input], [txt, chatbot])
|
259 |
-
txt.submit(user_send, [txt, state, image_input, video_input, pdf_input], [txt, chatbot])
|
260 |
|
261 |
-
|
262 |
-
|
263 |
-
demo.launch()
|
|
|
14 |
logging.basicConfig(level=logging.INFO)
|
15 |
logger = logging.getLogger(__name__)
|
16 |
|
17 |
+
# Check CUDA
|
18 |
logger.info(f"CUDA available: {torch.cuda.is_available()}")
|
19 |
if torch.cuda.is_available():
|
20 |
logger.info(f"GPU: {torch.cuda.get_device_name(0)}")
|
|
|
21 |
|
22 |
+
# BitsAndBytes config for quantized model loading
|
23 |
bnb_config = BitsAndBytesConfig(
|
24 |
load_in_4bit=True,
|
25 |
bnb_4bit_quant_type="nf4",
|
26 |
bnb_4bit_compute_dtype=torch.float16,
|
27 |
)
|
28 |
|
29 |
+
# Load Qwen model
|
30 |
try:
|
|
|
31 |
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Omni-3B", trust_remote_code=True)
|
32 |
model = AutoModelForCausalLM.from_pretrained(
|
33 |
"Qwen/Qwen2.5-Omni-3B",
|
|
|
35 |
quantization_config=bnb_config,
|
36 |
trust_remote_code=True
|
37 |
).eval()
|
38 |
+
logger.info("Qwen model loaded.")
|
39 |
except Exception as e:
|
40 |
+
logger.error(f"Failed to load Qwen: {e}")
|
41 |
+
model, tokenizer = None, None
|
|
|
42 |
|
43 |
+
# Load SentenceTransformer for RAG
|
44 |
try:
|
45 |
embed_model = SentenceTransformer('paraphrase-MiniLM-L3-v2')
|
46 |
+
logger.info("Embedding model loaded.")
|
47 |
except Exception as e:
|
48 |
+
logger.error(f"Failed to load embedding model: {e}")
|
49 |
embed_model = None
|
50 |
|
51 |
+
# Global index state
|
52 |
chunks = []
|
53 |
index = None
|
54 |
|
55 |
+
# PDF text chunking
|
56 |
def extract_chunks_from_pdf(pdf_path, chunk_size=1000, overlap=200):
|
57 |
try:
|
58 |
doc = fitz.open(pdf_path)
|
59 |
+
text = "".join([page.get_text() for page in doc])
|
|
|
|
|
60 |
return [text[i:i + chunk_size] for i in range(0, len(text), chunk_size - overlap)]
|
61 |
except Exception as e:
|
62 |
+
logger.error(f"PDF error: {e}")
|
63 |
+
return ["Error extracting content."]
|
64 |
|
65 |
+
# Build FAISS index
|
66 |
def build_faiss_index(chunks):
|
67 |
try:
|
|
|
|
|
68 |
embeddings = embed_model.encode(chunks, convert_to_numpy=True)
|
69 |
+
index = faiss.IndexFlatL2(embeddings.shape[1])
|
70 |
+
index.add(embeddings)
|
71 |
+
return index
|
|
|
72 |
except Exception as e:
|
73 |
logger.error(f"FAISS index error: {e}")
|
74 |
return None
|
75 |
|
76 |
+
# RAG retrieval
|
77 |
def rag_query(query, chunks, index, top_k=3):
|
|
|
|
|
78 |
try:
|
79 |
q_emb = embed_model.encode([query], convert_to_numpy=True)
|
80 |
D, I = index.search(q_emb, top_k)
|
81 |
return "\n\n".join([chunks[i] for i in I[0]])
|
82 |
except Exception as e:
|
83 |
logger.error(f"RAG query error: {e}")
|
84 |
+
return "Error retrieving context."
|
85 |
|
86 |
+
# Qwen chat
|
87 |
+
def chat_with_qwen(text, image=None):
|
88 |
if not model or not tokenizer:
|
89 |
+
return "Model not loaded."
|
|
|
90 |
try:
|
91 |
+
messages = [{"role": "user", "content": text}]
|
|
|
|
|
92 |
if image:
|
93 |
+
messages[0]["content"] = [{"image": image}, {"text": text}]
|
94 |
+
response, _ = model.chat(tokenizer, messages, history=None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
return response
|
96 |
except Exception as e:
|
97 |
logger.error(f"Chat error: {e}")
|
98 |
+
return f"Chat error: {e}"
|
99 |
|
100 |
+
# Extract representative frames
|
101 |
def extract_video_frames(video_path, max_frames=2):
|
102 |
try:
|
103 |
cap = cv2.VideoCapture(video_path)
|
104 |
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
105 |
+
frame_indices = [int(i * total_frames / max_frames) for i in range(max_frames)]
|
106 |
frames = []
|
107 |
+
for idx in frame_indices:
|
108 |
+
cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
|
109 |
+
success, frame = cap.read()
|
110 |
+
if success:
|
111 |
+
frames.append(frame)
|
|
|
|
|
|
|
|
|
112 |
cap.release()
|
113 |
return frames
|
114 |
except Exception as e:
|
115 |
+
logger.error(f"Frame extraction error: {e}")
|
116 |
return []
|
117 |
|
118 |
+
# Multimodal chat logic
|
119 |
def multimodal_chat(message, history, image=None, video=None, pdf=None):
|
120 |
global chunks, index
|
121 |
+
|
122 |
if not model:
|
123 |
+
return "Model not available."
|
124 |
|
125 |
try:
|
126 |
+
# PDF + question
|
127 |
+
if pdf and message:
|
128 |
+
pdf_path = pdf.name if hasattr(pdf, 'name') else None
|
129 |
+
if not pdf_path:
|
130 |
+
return "Invalid PDF input."
|
131 |
+
chunks = extract_chunks_from_pdf(pdf_path)
|
132 |
index = build_faiss_index(chunks)
|
133 |
if index:
|
134 |
context = rag_query(message, chunks, index)
|
135 |
+
user_prompt = f"Context:\n{context}\n\nQuestion: {message}"
|
136 |
+
return chat_with_qwen(user_prompt)
|
137 |
else:
|
138 |
+
return "Failed to process PDF."
|
|
|
139 |
|
140 |
+
# Image + question
|
141 |
+
if image and message:
|
142 |
+
return chat_with_qwen(message, image)
|
|
|
143 |
|
144 |
+
# Video + question
|
145 |
+
if video and message:
|
146 |
+
with tempfile.TemporaryDirectory() as temp_dir:
|
147 |
+
video_path = os.path.join(temp_dir, "video.mp4")
|
148 |
+
shutil.copy(video.name if hasattr(video, 'name') else video, video_path)
|
|
|
149 |
frames = extract_video_frames(video_path)
|
150 |
+
if not frames:
|
151 |
+
return "Could not extract video frames."
|
152 |
|
153 |
+
temp_img_path = os.path.join(temp_dir, "frame.jpg")
|
154 |
+
cv2.imwrite(temp_img_path, cv2.cvtColor(frames[0], cv2.COLOR_BGR2RGB))
|
155 |
+
return chat_with_qwen(message, temp_img_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
156 |
|
157 |
# Text only
|
158 |
if message:
|
159 |
return chat_with_qwen(message)
|
160 |
|
161 |
+
return "Please enter a question and optionally upload a file."
|
162 |
except Exception as e:
|
163 |
+
logger.error(f"Chat error: {e}")
|
164 |
+
return f"Error: {e}"
|
165 |
|
166 |
+
# Gradio UI
|
167 |
with gr.Blocks(css="""
|
168 |
+
body { background-color: #f3f6fc; }
|
169 |
+
.gradio-container { font-family: 'Segoe UI', sans-serif; }
|
|
|
|
|
|
|
|
|
170 |
h1 {
|
171 |
+
background: linear-gradient(to right, #667eea, #764ba2);
|
172 |
+
color: white !important;
|
173 |
+
padding: 1rem; border-radius: 12px; margin-bottom: 0.5rem;
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
}
|
175 |
.gr-box {
|
176 |
+
background-color: white; border-radius: 12px;
|
177 |
+
box-shadow: 0 0 10px rgba(0,0,0,0.05); padding: 16px;
|
|
|
|
|
178 |
}
|
179 |
+
footer { display: none !important; }
|
180 |
""") as demo:
|
181 |
+
|
182 |
+
gr.Markdown("""
|
183 |
+
<h1 style='text-align: center;'>Multimodal Chatbot powered by Qwen-2.5-Omni-3B</h1>
|
184 |
+
<p style='text-align: center;'>Ask your own questions with optional image, video, or PDF context.</p>
|
185 |
+
""")
|
186 |
|
187 |
chatbot = gr.Chatbot(show_label=False, height=450)
|
188 |
state = gr.State([])
|
189 |
|
190 |
with gr.Row():
|
191 |
+
txt = gr.Textbox(show_label=False, placeholder="Type your question...", scale=5)
|
192 |
send_btn = gr.Button("🚀 Send", scale=1)
|
193 |
|
194 |
with gr.Row():
|
|
|
198 |
|
199 |
def user_send(message, history, image, video, pdf):
|
200 |
if not message and not image and not video and not pdf:
|
201 |
+
return "", history, history
|
202 |
response = multimodal_chat(message, history, image, video, pdf)
|
203 |
history.append((message, response))
|
204 |
+
return "", history, history
|
205 |
|
206 |
+
send_btn.click(user_send, [txt, state, image_input, video_input, pdf_input], [txt, chatbot, state])
|
207 |
+
txt.submit(user_send, [txt, state, image_input, video_input, pdf_input], [txt, chatbot, state])
|
208 |
|
209 |
+
logger.info("Launching Gradio app")
|
210 |
+
demo.launch()
|
|