jimmy-yo commited on
Commit
2267ed1
·
verified ·
1 Parent(s): 96cf7ba

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +71 -60
app.py CHANGED
@@ -1,4 +1,5 @@
1
  import os
 
2
  import mimetypes
3
  from typing import List, Optional
4
 
@@ -7,7 +8,7 @@ import gradio as gr
7
  import numpy as np
8
  from PIL import Image
9
  import torch
10
- from transformers import AutoProcessor, AutoModelForVision2Seq
11
 
12
  # =========================
13
  # Config
@@ -19,11 +20,11 @@ MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
19
  DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
20
  DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
21
 
22
- MAX_VIDEO_SECONDS = 15 # hard limit for video duration we sample from
23
- MAX_FILE_MB = 60 # guard rail for very large files
24
- MAX_SIDE = 768 # resize longer side for images/frames
25
- MAX_FRAMES = 4 # how many frames to sample from video
26
- MAX_NEW_TOKENS = 256 # generation length
27
 
28
  # =========================
29
  # Load model & processor
@@ -36,17 +37,10 @@ model = AutoModelForVision2Seq.from_pretrained(
36
  trust_remote_code=True,
37
  ).to(DEVICE)
38
 
39
-
40
  # =========================
41
  # Helpers
42
  # =========================
43
  def get_path(file_obj) -> Optional[str]:
44
- """
45
- Gradio <File> may arrive as:
46
- - str (tmp path) -> return as-is
47
- - dict {"name": "..."} -> return ["name"] or ["path"]
48
- - SimpleNamespace with .name -> return .name
49
- """
50
  if file_obj is None:
51
  return None
52
  if isinstance(file_obj, str):
@@ -55,17 +49,14 @@ def get_path(file_obj) -> Optional[str]:
55
  return file_obj.get("name") or file_obj.get("path")
56
  return getattr(file_obj, "name", None)
57
 
58
-
59
  def is_image_file(path: str) -> bool:
60
  kind, _ = mimetypes.guess_type(path)
61
  return (kind or "").startswith("image/")
62
 
63
-
64
  def is_video_file(path: str) -> bool:
65
  kind, _ = mimetypes.guess_type(path)
66
  return (kind or "").startswith("video/")
67
 
68
-
69
  def resize_max_side(img: Image.Image, max_side: int = MAX_SIDE) -> Image.Image:
70
  w, h = img.size
71
  m = max(w, h)
@@ -74,27 +65,20 @@ def resize_max_side(img: Image.Image, max_side: int = MAX_SIDE) -> Image.Image:
74
  s = max_side / float(m)
75
  return img.resize((int(w * s), int(h * s)), Image.LANCZOS)
76
 
77
-
78
  def sample_video_frames(video_path: str, max_seconds=MAX_VIDEO_SECONDS, k=MAX_FRAMES) -> List[Image.Image]:
79
- """Take K evenly spaced frames within the first `max_seconds`."""
80
  cap = cv2.VideoCapture(video_path)
81
  if not cap.isOpened():
82
  return []
83
-
84
  fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
85
  total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
86
  duration = total / max(fps, 1)
87
-
88
- # Only consider frames within MAX_VIDEO_SECONDS (or all if shorter)
89
  limit_frames = int(min(duration, max_seconds) * fps) if total else 0
90
  nframes = min(limit_frames or total, total) if total > 0 else 0
91
  if nframes == 0:
92
  nframes = int(min(duration, max_seconds) * fps) or 1
93
-
94
  k = max(1, min(k, nframes))
95
  idxs = [int(i * nframes / k + nframes / (2 * k)) for i in range(k)]
96
  idxs = sorted(set(max(0, min(nframes - 1, x)) for x in idxs))
97
-
98
  frames: List[Image.Image] = []
99
  for target in idxs:
100
  cap.set(cv2.CAP_PROP_POS_FRAMES, target)
@@ -107,9 +91,7 @@ def sample_video_frames(video_path: str, max_seconds=MAX_VIDEO_SECONDS, k=MAX_FR
107
  cap.release()
108
  return frames
109
 
110
-
111
  def build_messages(prompt: str) -> list:
112
- """Qwen2-VL chat template expects <image> tag in user turn."""
113
  return [{
114
  "role": "user",
115
  "content": [
@@ -118,32 +100,33 @@ def build_messages(prompt: str) -> list:
118
  ]
119
  }]
120
 
121
-
122
  # =========================
123
- # Chat function
124
  # =========================
125
- def chat_fn(history, file_obj, prompt):
126
  history = history or []
127
- history.append([prompt, ""]) # add user bubble first
128
- yield history # show immediately
 
 
 
 
129
 
130
  path = get_path(file_obj)
131
  if not path or not os.path.exists(path):
132
  history[-1][1] = "⚠️ Please upload an image or a short video first."
133
- yield history
134
  return
135
 
136
- # Size guard
137
  try:
138
  if os.path.getsize(path) > MAX_FILE_MB * 1024 * 1024:
139
  history[-1][1] = f"⚠️ File is larger than {MAX_FILE_MB} MB."
140
- yield history
141
  return
142
  except Exception:
143
  pass
144
 
145
- # Collect images for the model
146
- images: List[Image.Image] = []
147
  try:
148
  if is_image_file(path):
149
  images = [resize_max_side(Image.open(path).convert("RGB"))]
@@ -151,48 +134,77 @@ def chat_fn(history, file_obj, prompt):
151
  images = sample_video_frames(path, MAX_VIDEO_SECONDS, MAX_FRAMES)
152
  if not images:
153
  history[-1][1] = "⚠️ Could not read frames from the video."
154
- yield history
155
  return
156
  else:
157
  history[-1][1] = "⚠️ Unsupported file type."
158
- yield history
159
  return
160
 
161
- # Build chat input
162
  messages = build_messages(prompt)
163
- chat_str = processor.apply_chat_template(
164
- messages, add_generation_prompt=True, tokenize=False
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
165
  )
166
 
167
- # Processor expects a LIST of images (even when it's 1)
168
- enc = processor(
169
- text=chat_str,
170
- images=images, # ALWAYS a list
171
- return_tensors="pt"
172
- ).to(model.device)
173
-
174
- # Inference (one-shot; more robust than streaming for Spaces CPU)
175
  with torch.no_grad():
176
  out_ids = model.generate(
177
  **enc,
178
  max_new_tokens=MAX_NEW_TOKENS,
179
  do_sample=False,
 
180
  )
181
 
182
- # Decode only the generated continuation
183
- cut = enc["input_ids"].shape[-1]
184
- text = processor.batch_decode(
185
- out_ids[:, cut:], skip_special_tokens=True
186
- )[0].strip()
 
 
 
187
 
188
- history[-1][1] = text if text else "(no output)"
189
- yield history
190
 
191
  except Exception as e:
192
- # Show clean error in the bubble
193
  history[-1][1] = f"⚠️ Error: {str(e)}"
194
- yield history
195
-
 
 
 
 
 
 
 
 
 
 
 
 
 
196
 
197
  # =========================
198
  # Gradio UI
@@ -223,7 +235,6 @@ with gr.Blocks(title=TITLE, css=CSS, theme=gr.themes.Soft()) as demo:
223
  )
224
  send = gr.Button("Submit", variant="primary")
225
 
226
- # Preview on file change
227
  def on_file(file_obj):
228
  path = get_path(file_obj)
229
  if not path or not os.path.exists(path):
@@ -236,7 +247,7 @@ with gr.Blocks(title=TITLE, css=CSS, theme=gr.themes.Soft()) as demo:
236
  return None
237
 
238
  uploader.change(on_file, inputs=[uploader], outputs=[preview])
239
- send.click(chat_fn, inputs=[chat, uploader, prompt_in], outputs=[chat])
240
 
241
  if __name__ == "__main__":
242
  demo.queue().launch()
 
1
  import os
2
+ import time
3
  import mimetypes
4
  from typing import List, Optional
5
 
 
8
  import numpy as np
9
  from PIL import Image
10
  import torch
11
+ from transformers import AutoProcessor, AutoModelForVision2Seq, TextStreamer
12
 
13
  # =========================
14
  # Config
 
20
  DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
21
  DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
22
 
23
+ MAX_VIDEO_SECONDS = 15
24
+ MAX_FILE_MB = 60
25
+ MAX_SIDE = 768
26
+ MAX_FRAMES = 4
27
+ MAX_NEW_TOKENS = 256
28
 
29
  # =========================
30
  # Load model & processor
 
37
  trust_remote_code=True,
38
  ).to(DEVICE)
39
 
 
40
  # =========================
41
  # Helpers
42
  # =========================
43
  def get_path(file_obj) -> Optional[str]:
 
 
 
 
 
 
44
  if file_obj is None:
45
  return None
46
  if isinstance(file_obj, str):
 
49
  return file_obj.get("name") or file_obj.get("path")
50
  return getattr(file_obj, "name", None)
51
 
 
52
  def is_image_file(path: str) -> bool:
53
  kind, _ = mimetypes.guess_type(path)
54
  return (kind or "").startswith("image/")
55
 
 
56
  def is_video_file(path: str) -> bool:
57
  kind, _ = mimetypes.guess_type(path)
58
  return (kind or "").startswith("video/")
59
 
 
60
  def resize_max_side(img: Image.Image, max_side: int = MAX_SIDE) -> Image.Image:
61
  w, h = img.size
62
  m = max(w, h)
 
65
  s = max_side / float(m)
66
  return img.resize((int(w * s), int(h * s)), Image.LANCZOS)
67
 
 
68
  def sample_video_frames(video_path: str, max_seconds=MAX_VIDEO_SECONDS, k=MAX_FRAMES) -> List[Image.Image]:
 
69
  cap = cv2.VideoCapture(video_path)
70
  if not cap.isOpened():
71
  return []
 
72
  fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
73
  total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
74
  duration = total / max(fps, 1)
 
 
75
  limit_frames = int(min(duration, max_seconds) * fps) if total else 0
76
  nframes = min(limit_frames or total, total) if total > 0 else 0
77
  if nframes == 0:
78
  nframes = int(min(duration, max_seconds) * fps) or 1
 
79
  k = max(1, min(k, nframes))
80
  idxs = [int(i * nframes / k + nframes / (2 * k)) for i in range(k)]
81
  idxs = sorted(set(max(0, min(nframes - 1, x)) for x in idxs))
 
82
  frames: List[Image.Image] = []
83
  for target in idxs:
84
  cap.set(cv2.CAP_PROP_POS_FRAMES, target)
 
91
  cap.release()
92
  return frames
93
 
 
94
  def build_messages(prompt: str) -> list:
 
95
  return [{
96
  "role": "user",
97
  "content": [
 
100
  ]
101
  }]
102
 
 
103
  # =========================
104
+ # Chat (streaming)
105
  # =========================
106
+ def chat_stream(history, file_obj, prompt):
107
  history = history or []
108
+ # 1) Add user turn
109
+ history.append([prompt, ""])
110
+ # 2) Add assistant placeholder
111
+ history.append(["", "⏳ processing…"])
112
+ # Immediately show placeholder and disable inputs
113
+ yield history, gr.update(interactive=False), gr.update(interactive=False)
114
 
115
  path = get_path(file_obj)
116
  if not path or not os.path.exists(path):
117
  history[-1][1] = "⚠️ Please upload an image or a short video first."
118
+ yield history, gr.update(interactive=True), gr.update(interactive=True)
119
  return
120
 
 
121
  try:
122
  if os.path.getsize(path) > MAX_FILE_MB * 1024 * 1024:
123
  history[-1][1] = f"⚠️ File is larger than {MAX_FILE_MB} MB."
124
+ yield history, gr.update(interactive=True), gr.update(interactive=True)
125
  return
126
  except Exception:
127
  pass
128
 
129
+ # Prepare images list
 
130
  try:
131
  if is_image_file(path):
132
  images = [resize_max_side(Image.open(path).convert("RGB"))]
 
134
  images = sample_video_frames(path, MAX_VIDEO_SECONDS, MAX_FRAMES)
135
  if not images:
136
  history[-1][1] = "⚠️ Could not read frames from the video."
137
+ yield history, gr.update(interactive=True), gr.update(interactive=True)
138
  return
139
  else:
140
  history[-1][1] = "⚠️ Unsupported file type."
141
+ yield history, gr.update(interactive=True), gr.update(interactive=True)
142
  return
143
 
 
144
  messages = build_messages(prompt)
145
+ chat_str = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
146
+ enc = processor(text=chat_str, images=images, return_tensors="pt").to(model.device)
147
+
148
+ # Streaming callback
149
+ acc = []
150
+ last_push = time.time()
151
+
152
+ def on_text(chunk: str):
153
+ nonlocal acc, last_push, history
154
+ acc.append(chunk)
155
+ # throttle UI updates to avoid spamming the frontend
156
+ if time.time() - last_push > 0.06:
157
+ history[-1][1] = "".join(acc)
158
+ last_push = time.time()
159
+ # Keep inputs disabled while streaming
160
+ yield_state = (history, gr.update(interactive=False), gr.update(interactive=False))
161
+ # Gradio requires generator; hack by storing to function attr
162
+ chat_stream._yield = yield_state # type: ignore
163
+
164
+ streamer = TextStreamer(
165
+ processor.tokenizer,
166
+ skip_prompt=True,
167
+ skip_special_tokens=True,
168
+ callback=on_text,
169
  )
170
 
171
+ # Kick off generation
 
 
 
 
 
 
 
172
  with torch.no_grad():
173
  out_ids = model.generate(
174
  **enc,
175
  max_new_tokens=MAX_NEW_TOKENS,
176
  do_sample=False,
177
+ streamer=streamer,
178
  )
179
 
180
+ # Flush final text
181
+ if acc:
182
+ history[-1][1] = "".join(acc).strip()
183
+ else:
184
+ # Fallback (non-stream)
185
+ cut = enc["input_ids"].shape[-1]
186
+ text = processor.batch_decode(out_ids[:, cut:], skip_special_tokens=True)[0].strip()
187
+ history[-1][1] = text if text else "(no output)"
188
 
189
+ yield history, gr.update(interactive=True), gr.update(interactive=True)
 
190
 
191
  except Exception as e:
 
192
  history[-1][1] = f"⚠️ Error: {str(e)}"
193
+ yield history, gr.update(interactive=True), gr.update(interactive=True)
194
+
195
+ # A tiny wrapper to surface mid-stream yields from callback
196
+ def _pump_stream(*args, **kwargs):
197
+ gen = chat_stream(*args, **kwargs)
198
+ while True:
199
+ # Check if callback parked a mid-step yield
200
+ ys = getattr(chat_stream, "_yield", None)
201
+ if ys is not None:
202
+ setattr(chat_stream, "_yield", None)
203
+ yield ys
204
+ try:
205
+ yield next(gen)
206
+ except StopIteration:
207
+ break
208
 
209
  # =========================
210
  # Gradio UI
 
235
  )
236
  send = gr.Button("Submit", variant="primary")
237
 
 
238
  def on_file(file_obj):
239
  path = get_path(file_obj)
240
  if not path or not os.path.exists(path):
 
247
  return None
248
 
249
  uploader.change(on_file, inputs=[uploader], outputs=[preview])
250
+ send.click(_pump_stream, inputs=[chat, uploader, prompt_in], outputs=[chat, send, prompt_in])
251
 
252
  if __name__ == "__main__":
253
  demo.queue().launch()