Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -14,9 +14,7 @@ from transformers import (
|
|
14 |
)
|
15 |
from transformers import Qwen2_5_VLForConditionalGeneration
|
16 |
|
17 |
-
# ---------------------------
|
18 |
# Helper Functions
|
19 |
-
# ---------------------------
|
20 |
def progress_bar_html(label: str, primary_color: str = "#4B0082", secondary_color: str = "#9370DB") -> str:
|
21 |
"""
|
22 |
Returns an HTML snippet for a thin animated progress bar with a label.
|
@@ -49,7 +47,6 @@ def downsample_video(video_path):
|
|
49 |
if total_frames <= 0 or fps <= 0:
|
50 |
vidcap.release()
|
51 |
return frames
|
52 |
-
# Determine 10 evenly spaced frame indices.
|
53 |
frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
|
54 |
for i in frame_indices:
|
55 |
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
@@ -63,8 +60,7 @@ def downsample_video(video_path):
|
|
63 |
return frames
|
64 |
|
65 |
# Model and Processor Setup
|
66 |
-
|
67 |
-
QV_MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct" # [or] prithivMLmods/Qwen2-VL-OCR2-2B-Instruct
|
68 |
qwen_processor = AutoProcessor.from_pretrained(QV_MODEL_ID, trust_remote_code=True)
|
69 |
qwen_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
70 |
QV_MODEL_ID,
|
@@ -72,7 +68,6 @@ qwen_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
|
72 |
torch_dtype=torch.float16
|
73 |
).to("cuda").eval()
|
74 |
|
75 |
-
# RolmOCR branch (@RolmOCR)
|
76 |
ROLMOCR_MODEL_ID = "reducto/RolmOCR"
|
77 |
rolmocr_processor = AutoProcessor.from_pretrained(ROLMOCR_MODEL_ID, trust_remote_code=True)
|
78 |
rolmocr_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
@@ -83,111 +78,62 @@ rolmocr_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
|
83 |
|
84 |
# Main Inference Function
|
85 |
@spaces.GPU
|
86 |
-
def model_inference(input_dict, history):
|
87 |
text = input_dict["text"].strip()
|
88 |
files = input_dict.get("files", [])
|
89 |
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
|
|
98 |
if not frames:
|
99 |
yield "Error: Could not extract frames from the video."
|
100 |
return
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
content_list.append({"type": "text", "text": f"Frame {timestamp}:"})
|
105 |
-
content_list.append({"type": "image", "image": image})
|
106 |
-
messages = [{"role": "user", "content": content_list}]
|
107 |
-
# For video, extract images only.
|
108 |
-
video_images = [image for image, _ in frames]
|
109 |
-
prompt_full = rolmocr_processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
110 |
-
inputs = rolmocr_processor(
|
111 |
-
text=[prompt_full],
|
112 |
-
images=video_images,
|
113 |
-
return_tensors="pt",
|
114 |
-
padding=True,
|
115 |
-
).to("cuda")
|
116 |
else:
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
images = []
|
124 |
-
if text_prompt == "" and not images:
|
125 |
-
yield "Error: Please input a text query and/or provide an image for the @RolmOCR feature."
|
126 |
return
|
127 |
-
messages = [{
|
128 |
-
"role": "user",
|
129 |
-
"content": [
|
130 |
-
*[{"type": "image", "image": image} for image in images],
|
131 |
-
{"type": "text", "text": text_prompt},
|
132 |
-
],
|
133 |
-
}]
|
134 |
-
prompt_full = rolmocr_processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
135 |
-
inputs = rolmocr_processor(
|
136 |
-
text=[prompt_full],
|
137 |
-
images=images if images else None,
|
138 |
-
return_tensors="pt",
|
139 |
-
padding=True,
|
140 |
-
).to("cuda")
|
141 |
-
streamer = TextIteratorStreamer(rolmocr_processor, skip_prompt=True, skip_special_tokens=True)
|
142 |
-
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
143 |
-
thread = Thread(target=rolmocr_model.generate, kwargs=generation_kwargs)
|
144 |
-
thread.start()
|
145 |
-
buffer = ""
|
146 |
-
# Use a different color scheme for RolmOCR (purple-themed).
|
147 |
-
yield progress_bar_html("Processing with Qwen2.5VL (RolmOCR)")
|
148 |
-
for new_text in streamer:
|
149 |
-
buffer += new_text
|
150 |
-
buffer = buffer.replace("<|im_end|>", "")
|
151 |
-
time.sleep(0.01)
|
152 |
-
yield buffer
|
153 |
-
return
|
154 |
|
155 |
-
#
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
images = []
|
163 |
-
|
164 |
-
if text == "" and not images:
|
165 |
-
yield "Error: Please input a text query and optionally image(s)."
|
166 |
-
return
|
167 |
-
if text == "" and images:
|
168 |
-
yield "Error: Please input a text query along with the image(s)."
|
169 |
-
return
|
170 |
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
inputs = qwen_processor(
|
180 |
text=[prompt_full],
|
181 |
-
images=
|
182 |
return_tensors="pt",
|
183 |
padding=True,
|
184 |
).to("cuda")
|
185 |
-
|
|
|
186 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
187 |
-
thread = Thread(target=
|
188 |
thread.start()
|
189 |
buffer = ""
|
190 |
-
yield progress_bar_html("Processing with
|
191 |
for new_text in streamer:
|
192 |
buffer += new_text
|
193 |
buffer = buffer.replace("<|im_end|>", "")
|
@@ -196,25 +142,26 @@ def model_inference(input_dict, history):
|
|
196 |
|
197 |
# Gradio Interface
|
198 |
examples = [
|
199 |
-
[{"text": "
|
200 |
-
[{"text": "
|
201 |
-
[{"text": "
|
202 |
[{"text": "Extract as JSON table from the table", "files": ["examples/4.jpg"]}],
|
203 |
]
|
204 |
|
205 |
demo = gr.ChatInterface(
|
206 |
fn=model_inference,
|
207 |
-
description="# **Multimodal OCR
|
208 |
examples=examples,
|
209 |
textbox=gr.MultimodalTextbox(
|
210 |
-
label="Query Input",
|
211 |
-
file_types=["image", "video"],
|
212 |
-
file_count="multiple",
|
213 |
-
placeholder="
|
214 |
),
|
215 |
stop_btn="Stop Generation",
|
216 |
multimodal=True,
|
217 |
cache_examples=False,
|
|
|
218 |
)
|
219 |
|
220 |
demo.launch(debug=True)
|
|
|
14 |
)
|
15 |
from transformers import Qwen2_5_VLForConditionalGeneration
|
16 |
|
|
|
17 |
# Helper Functions
|
|
|
18 |
def progress_bar_html(label: str, primary_color: str = "#4B0082", secondary_color: str = "#9370DB") -> str:
|
19 |
"""
|
20 |
Returns an HTML snippet for a thin animated progress bar with a label.
|
|
|
47 |
if total_frames <= 0 or fps <= 0:
|
48 |
vidcap.release()
|
49 |
return frames
|
|
|
50 |
frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
|
51 |
for i in frame_indices:
|
52 |
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
|
|
60 |
return frames
|
61 |
|
62 |
# Model and Processor Setup
|
63 |
+
QV_MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
|
|
64 |
qwen_processor = AutoProcessor.from_pretrained(QV_MODEL_ID, trust_remote_code=True)
|
65 |
qwen_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
66 |
QV_MODEL_ID,
|
|
|
68 |
torch_dtype=torch.float16
|
69 |
).to("cuda").eval()
|
70 |
|
|
|
71 |
ROLMOCR_MODEL_ID = "reducto/RolmOCR"
|
72 |
rolmocr_processor = AutoProcessor.from_pretrained(ROLMOCR_MODEL_ID, trust_remote_code=True)
|
73 |
rolmocr_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
|
|
78 |
|
79 |
# Main Inference Function
|
80 |
@spaces.GPU
|
81 |
+
def model_inference(input_dict, history, use_rolmocr=False):
|
82 |
text = input_dict["text"].strip()
|
83 |
files = input_dict.get("files", [])
|
84 |
|
85 |
+
if not text and not files:
|
86 |
+
yield "Error: Please input a text query or provide files (images or videos)."
|
87 |
+
return
|
88 |
+
|
89 |
+
# Process files: images and videos
|
90 |
+
image_list = []
|
91 |
+
for idx, file in enumerate(files):
|
92 |
+
if file.lower().endswith((".mp4", ".avi", ".mov")):
|
93 |
+
frames = downsample_video(file)
|
94 |
if not frames:
|
95 |
yield "Error: Could not extract frames from the video."
|
96 |
return
|
97 |
+
for frame, timestamp in frames:
|
98 |
+
label = f"Video {idx+1} Frame {timestamp}:"
|
99 |
+
image_list.append((label, frame))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
else:
|
101 |
+
try:
|
102 |
+
img = load_image(file)
|
103 |
+
label = f"Image {idx+1}:"
|
104 |
+
image_list.append((label, img))
|
105 |
+
except Exception as e:
|
106 |
+
yield f"Error loading image: {str(e)}"
|
|
|
|
|
|
|
107 |
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
109 |
+
# Build content list
|
110 |
+
content = [{"type": "text", "text": text}]
|
111 |
+
for label, img in image_list:
|
112 |
+
content.append({"type": "text", "text": label})
|
113 |
+
content.append({"type": "image", "image": img})
|
114 |
+
|
115 |
+
messages = [{"role": "user", "content": content}]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
|
117 |
+
# Select processor and model
|
118 |
+
processor = rolmocr_processor if use_rolmocr else qwen_processor
|
119 |
+
model = rolmocr_model if use_rolmocr else qwen_model
|
120 |
+
model_name = "RolmOCR" if use_rolmocr else "Qwen2VL OCR"
|
121 |
+
|
122 |
+
prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
123 |
+
all_images = [item["image"] for item in content if item["type"] == "image"]
|
124 |
+
inputs = processor(
|
|
|
125 |
text=[prompt_full],
|
126 |
+
images=all_images if all_images else None,
|
127 |
return_tensors="pt",
|
128 |
padding=True,
|
129 |
).to("cuda")
|
130 |
+
|
131 |
+
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
132 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
133 |
+
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
134 |
thread.start()
|
135 |
buffer = ""
|
136 |
+
yield progress_bar_html(f"Processing with {model_name}")
|
137 |
for new_text in streamer:
|
138 |
buffer += new_text
|
139 |
buffer = buffer.replace("<|im_end|>", "")
|
|
|
142 |
|
143 |
# Gradio Interface
|
144 |
examples = [
|
145 |
+
[{"text": "OCR the Text in the Image", "files": ["rolm/1.jpeg"]}],
|
146 |
+
[{"text": "Explain the Ad in Detail", "files": ["examples/videoplayback.mp4"]}],
|
147 |
+
[{"text": "OCR the Image", "files": ["rolm/3.jpeg"]}],
|
148 |
[{"text": "Extract as JSON table from the table", "files": ["examples/4.jpg"]}],
|
149 |
]
|
150 |
|
151 |
demo = gr.ChatInterface(
|
152 |
fn=model_inference,
|
153 |
+
description="# **Multimodal OCR with Model Selection**",
|
154 |
examples=examples,
|
155 |
textbox=gr.MultimodalTextbox(
|
156 |
+
label="Query Input",
|
157 |
+
file_types=["image", "video"],
|
158 |
+
file_count="multiple",
|
159 |
+
placeholder="Input your query and optionally upload image(s) or video(s). Select the model using the checkbox."
|
160 |
),
|
161 |
stop_btn="Stop Generation",
|
162 |
multimodal=True,
|
163 |
cache_examples=False,
|
164 |
+
additional_inputs=[gr.Checkbox(label="Use RolmOCR", value=False, info="Check to use RolmOCR, uncheck to use Qwen2VL OCR")],
|
165 |
)
|
166 |
|
167 |
demo.launch(debug=True)
|