Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -12,6 +12,7 @@ import torch
|
|
12 |
import numpy as np
|
13 |
from PIL import Image
|
14 |
import edge_tts
|
|
|
15 |
|
16 |
from transformers import (
|
17 |
AutoModelForCausalLM,
|
@@ -149,6 +150,28 @@ def progress_bar_html(label: str) -> str:
|
|
149 |
</style>
|
150 |
'''
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
@spaces.GPU(duration=60, enable_queue=True)
|
153 |
def generate_image_fn(
|
154 |
prompt: str,
|
@@ -213,14 +236,16 @@ def generate(
|
|
213 |
Special commands:
|
214 |
- "@tts1" or "@tts2": triggers text-to-speech.
|
215 |
- "@image": triggers image generation using the SDXL pipeline.
|
|
|
216 |
"""
|
217 |
text = input_dict["text"]
|
218 |
files = input_dict.get("files", [])
|
|
|
219 |
|
220 |
-
|
|
|
221 |
# Remove the "@image" tag and use the rest as prompt
|
222 |
prompt = text[len("@image"):].strip()
|
223 |
-
# Show animated progress bar for image generation
|
224 |
yield progress_bar_html("Generating Image")
|
225 |
image_paths, used_seed = generate_image_fn(
|
226 |
prompt=prompt,
|
@@ -235,10 +260,57 @@ def generate(
|
|
235 |
use_resolution_binning=True,
|
236 |
num_images=1,
|
237 |
)
|
238 |
-
# Once done, yield the generated image
|
239 |
yield gr.Image(image_paths[0])
|
240 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
|
|
|
242 |
tts_prefix = "@tts"
|
243 |
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 3))
|
244 |
voice_index = next((i for i in range(1, 3) if text.strip().lower().startswith(f"{tts_prefix}{i}")), None)
|
@@ -246,11 +318,9 @@ def generate(
|
|
246 |
if is_tts and voice_index:
|
247 |
voice = TTS_VOICES[voice_index - 1]
|
248 |
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
249 |
-
# Clear previous chat history for a fresh TTS request.
|
250 |
conversation = [{"role": "user", "content": text}]
|
251 |
else:
|
252 |
voice = None
|
253 |
-
# Remove any stray @tts tags and build the conversation history.
|
254 |
text = text.replace(tts_prefix, "").strip()
|
255 |
conversation = clean_chat_history(chat_history)
|
256 |
conversation.append({"role": "user", "content": text})
|
@@ -269,15 +339,13 @@ def generate(
|
|
269 |
{"type": "text", "text": text},
|
270 |
]
|
271 |
}]
|
272 |
-
|
273 |
-
inputs = processor(text=[
|
274 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
275 |
generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
|
276 |
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
277 |
thread.start()
|
278 |
-
|
279 |
buffer = ""
|
280 |
-
# Show animated progress bar for multimodal generation
|
281 |
yield progress_bar_html("Thinking...")
|
282 |
for new_text in streamer:
|
283 |
buffer += new_text
|
@@ -304,18 +372,13 @@ def generate(
|
|
304 |
}
|
305 |
t = Thread(target=model.generate, kwargs=generation_kwargs)
|
306 |
t.start()
|
307 |
-
|
308 |
outputs = []
|
309 |
-
|
310 |
-
yield progress_bar_html("Thinking...")
|
311 |
for new_text in streamer:
|
312 |
outputs.append(new_text)
|
313 |
yield "".join(outputs)
|
314 |
-
|
315 |
final_response = "".join(outputs)
|
316 |
yield final_response
|
317 |
-
|
318 |
-
# If TTS was requested, convert the final response to speech.
|
319 |
if is_tts and voice:
|
320 |
output_file = asyncio.run(text_to_speech(final_response, voice))
|
321 |
yield gr.Audio(output_file, autoplay=True)
|
@@ -330,6 +393,7 @@ demo = gr.ChatInterface(
|
|
330 |
gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2),
|
331 |
],
|
332 |
examples=[
|
|
|
333 |
["@image Chocolate dripping from a donut against a yellow background, in the style of brocore, hyper-realistic"],
|
334 |
["Python Program for Array Rotation"],
|
335 |
["@tts1 Who is Nikola Tesla, and why did he die?"],
|
@@ -342,7 +406,7 @@ demo = gr.ChatInterface(
|
|
342 |
description=DESCRIPTION,
|
343 |
css=css,
|
344 |
fill_height=True,
|
345 |
-
textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image"], file_count="multiple",
|
346 |
stop_btn="Stop Generation",
|
347 |
multimodal=True,
|
348 |
)
|
|
|
12 |
import numpy as np
|
13 |
from PIL import Image
|
14 |
import edge_tts
|
15 |
+
import cv2
|
16 |
|
17 |
from transformers import (
|
18 |
AutoModelForCausalLM,
|
|
|
150 |
</style>
|
151 |
'''
|
152 |
|
153 |
+
def downsample_video(video_path):
|
154 |
+
"""
|
155 |
+
Downsamples the video to 10 evenly spaced frames.
|
156 |
+
Each frame is returned as a PIL image along with its timestamp.
|
157 |
+
"""
|
158 |
+
vidcap = cv2.VideoCapture(video_path)
|
159 |
+
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
160 |
+
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
161 |
+
frames = []
|
162 |
+
# Sample 10 evenly spaced frames.
|
163 |
+
frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
|
164 |
+
for i in frame_indices:
|
165 |
+
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
166 |
+
success, image = vidcap.read()
|
167 |
+
if success:
|
168 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB
|
169 |
+
pil_image = Image.fromarray(image)
|
170 |
+
timestamp = round(i / fps, 2)
|
171 |
+
frames.append((pil_image, timestamp))
|
172 |
+
vidcap.release()
|
173 |
+
return frames
|
174 |
+
|
175 |
@spaces.GPU(duration=60, enable_queue=True)
|
176 |
def generate_image_fn(
|
177 |
prompt: str,
|
|
|
236 |
Special commands:
|
237 |
- "@tts1" or "@tts2": triggers text-to-speech.
|
238 |
- "@image": triggers image generation using the SDXL pipeline.
|
239 |
+
- "@qwen2vl-video": triggers video processing using Qwen2VL.
|
240 |
"""
|
241 |
text = input_dict["text"]
|
242 |
files = input_dict.get("files", [])
|
243 |
+
lower_text = text.strip().lower()
|
244 |
|
245 |
+
# Branch for image generation.
|
246 |
+
if lower_text.startswith("@image"):
|
247 |
# Remove the "@image" tag and use the rest as prompt
|
248 |
prompt = text[len("@image"):].strip()
|
|
|
249 |
yield progress_bar_html("Generating Image")
|
250 |
image_paths, used_seed = generate_image_fn(
|
251 |
prompt=prompt,
|
|
|
260 |
use_resolution_binning=True,
|
261 |
num_images=1,
|
262 |
)
|
|
|
263 |
yield gr.Image(image_paths[0])
|
264 |
+
return
|
265 |
+
|
266 |
+
# New branch for video processing with Qwen2VL.
|
267 |
+
if lower_text.startswith("@qwen2vl-video"):
|
268 |
+
prompt = text[len("@qwen2vl-video"):].strip()
|
269 |
+
if files:
|
270 |
+
# Assume the first file is a video.
|
271 |
+
video_path = files[0]
|
272 |
+
frames = downsample_video(video_path)
|
273 |
+
messages = [
|
274 |
+
{"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
|
275 |
+
{"role": "user", "content": [{"type": "text", "text": prompt}]}
|
276 |
+
]
|
277 |
+
# Append each frame with its timestamp.
|
278 |
+
for frame in frames:
|
279 |
+
image, timestamp = frame
|
280 |
+
image_path = f"video_frame_{uuid.uuid4().hex}.png"
|
281 |
+
image.save(image_path)
|
282 |
+
messages[1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
|
283 |
+
messages[1]["content"].append({"type": "image", "url": image_path})
|
284 |
+
else:
|
285 |
+
messages = [
|
286 |
+
{"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
|
287 |
+
{"role": "user", "content": [{"type": "text", "text": prompt}]}
|
288 |
+
]
|
289 |
+
inputs = processor.apply_chat_template(
|
290 |
+
messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
|
291 |
+
).to("cuda")
|
292 |
+
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
293 |
+
generation_kwargs = {
|
294 |
+
**inputs,
|
295 |
+
"streamer": streamer,
|
296 |
+
"max_new_tokens": max_new_tokens,
|
297 |
+
"do_sample": True,
|
298 |
+
"temperature": temperature,
|
299 |
+
"top_p": top_p,
|
300 |
+
"top_k": top_k,
|
301 |
+
"repetition_penalty": repetition_penalty,
|
302 |
+
}
|
303 |
+
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
304 |
+
thread.start()
|
305 |
+
buffer = ""
|
306 |
+
yield progress_bar_html("Processing video with Qwen2VL")
|
307 |
+
for new_text in streamer:
|
308 |
+
buffer += new_text
|
309 |
+
time.sleep(0.01)
|
310 |
+
yield buffer
|
311 |
+
return
|
312 |
|
313 |
+
# Determine if TTS is requested.
|
314 |
tts_prefix = "@tts"
|
315 |
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 3))
|
316 |
voice_index = next((i for i in range(1, 3) if text.strip().lower().startswith(f"{tts_prefix}{i}")), None)
|
|
|
318 |
if is_tts and voice_index:
|
319 |
voice = TTS_VOICES[voice_index - 1]
|
320 |
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
|
|
321 |
conversation = [{"role": "user", "content": text}]
|
322 |
else:
|
323 |
voice = None
|
|
|
324 |
text = text.replace(tts_prefix, "").strip()
|
325 |
conversation = clean_chat_history(chat_history)
|
326 |
conversation.append({"role": "user", "content": text})
|
|
|
339 |
{"type": "text", "text": text},
|
340 |
]
|
341 |
}]
|
342 |
+
prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
343 |
+
inputs = processor(text=[prompt_full], images=images, return_tensors="pt", padding=True).to("cuda")
|
344 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
345 |
generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
|
346 |
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
347 |
thread.start()
|
|
|
348 |
buffer = ""
|
|
|
349 |
yield progress_bar_html("Thinking...")
|
350 |
for new_text in streamer:
|
351 |
buffer += new_text
|
|
|
372 |
}
|
373 |
t = Thread(target=model.generate, kwargs=generation_kwargs)
|
374 |
t.start()
|
|
|
375 |
outputs = []
|
376 |
+
yield progress_bar_html("Processing with Qwen2VL Ocr")
|
|
|
377 |
for new_text in streamer:
|
378 |
outputs.append(new_text)
|
379 |
yield "".join(outputs)
|
|
|
380 |
final_response = "".join(outputs)
|
381 |
yield final_response
|
|
|
|
|
382 |
if is_tts and voice:
|
383 |
output_file = asyncio.run(text_to_speech(final_response, voice))
|
384 |
yield gr.Audio(output_file, autoplay=True)
|
|
|
393 |
gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2),
|
394 |
],
|
395 |
examples=[
|
396 |
+
[{"text": "@gemma3-4b-video Summarize the events in this video", "files": ["examples/sky.mp4"]}],
|
397 |
["@image Chocolate dripping from a donut against a yellow background, in the style of brocore, hyper-realistic"],
|
398 |
["Python Program for Array Rotation"],
|
399 |
["@tts1 Who is Nikola Tesla, and why did he die?"],
|
|
|
406 |
description=DESCRIPTION,
|
407 |
css=css,
|
408 |
fill_height=True,
|
409 |
+
textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image", "video"], file_count="multiple", placeholder=" @tts1, @tts2-voices, @image for image gen, @qwen2vl-video for video, default [text, vision]"),
|
410 |
stop_btn="Stop Generation",
|
411 |
multimodal=True,
|
412 |
)
|