Spaces:
Running
Running
updated device management
Browse files
app.py
CHANGED
@@ -17,7 +17,14 @@ from datetime import datetime
|
|
17 |
|
18 |
@st.cache_resource
|
19 |
def load_model(model_name):
|
20 |
-
"""Load the selected model and processor
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
try:
|
22 |
if model_name == "Donut":
|
23 |
processor = DonutProcessor.from_pretrained("naver-clova-ix/donut-base")
|
@@ -27,63 +34,98 @@ def load_model(model_name):
|
|
27 |
model.config.pad_token_id = processor.tokenizer.pad_token_id
|
28 |
model.config.vocab_size = len(processor.tokenizer)
|
29 |
|
|
|
|
|
30 |
elif model_name == "LayoutLMv3":
|
31 |
processor = LayoutLMv3Processor.from_pretrained("microsoft/layoutlmv3-base")
|
32 |
model = LayoutLMv3ForSequenceClassification.from_pretrained("microsoft/layoutlmv3-base")
|
33 |
|
|
|
|
|
34 |
elif model_name == "OmniParser":
|
35 |
# Load YOLO model for icon detection
|
36 |
-
yolo_model = YOLO("microsoft/OmniParser
|
37 |
-
|
38 |
-
processor
|
39 |
-
|
40 |
-
"microsoft/
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
trust_remote_code=True
|
42 |
)
|
43 |
|
44 |
return {
|
45 |
'yolo': yolo_model,
|
46 |
'processor': processor,
|
47 |
-
'model':
|
48 |
}
|
49 |
|
50 |
-
|
51 |
-
|
|
|
52 |
except Exception as e:
|
53 |
st.error(f"Error loading model {model_name}: {str(e)}")
|
54 |
-
return None
|
55 |
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
try:
|
|
|
|
|
|
|
59 |
if model_name == "OmniParser":
|
60 |
-
#
|
|
|
|
|
|
|
|
|
61 |
temp_path = "temp_image.png"
|
62 |
image.save(temp_path)
|
63 |
|
64 |
-
# Configure box detection parameters
|
65 |
-
box_threshold = 0.05
|
66 |
-
iou_threshold = 0.1
|
67 |
-
|
68 |
# Run YOLO detection
|
69 |
-
yolo_results =
|
70 |
temp_path,
|
71 |
conf=box_threshold,
|
72 |
iou=iou_threshold
|
73 |
)
|
74 |
|
75 |
-
# Process detections
|
76 |
results = []
|
77 |
for det in yolo_results[0].boxes.data:
|
78 |
x1, y1, x2, y2, conf, cls = det
|
79 |
|
80 |
# Get region of interest
|
81 |
-
roi = image.crop((x1, y1, x2, y2))
|
82 |
|
83 |
# Generate caption using the model
|
84 |
-
inputs = processor(
|
85 |
-
|
86 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
results.append({
|
89 |
"bbox": [float(x) for x in [x1, y1, x2, y2]],
|
@@ -92,31 +134,40 @@ def analyze_document(image, model_name, model, processor):
|
|
92 |
"caption": caption
|
93 |
})
|
94 |
|
|
|
|
|
|
|
|
|
95 |
return {
|
96 |
"detected_elements": len(results),
|
97 |
"elements": results
|
98 |
}
|
99 |
|
100 |
elif model_name == "Donut":
|
101 |
-
#
|
102 |
-
pixel_values = processor(image, return_tensors="pt").pixel_values
|
|
|
103 |
task_prompt = "<s_cord>analyze the document and extract information</s_cord>"
|
104 |
-
decoder_input_ids = processor.tokenizer(
|
|
|
|
|
|
|
|
|
105 |
|
106 |
-
outputs = model.generate(
|
107 |
pixel_values,
|
108 |
decoder_input_ids=decoder_input_ids,
|
109 |
max_length=512,
|
110 |
early_stopping=True,
|
111 |
-
pad_token_id=processor.tokenizer.pad_token_id,
|
112 |
-
eos_token_id=processor.tokenizer.eos_token_id,
|
113 |
use_cache=True,
|
114 |
num_beams=4,
|
115 |
-
bad_words_ids=[[processor.tokenizer.unk_token_id]],
|
116 |
return_dict_in_generate=True
|
117 |
)
|
118 |
|
119 |
-
sequence = processor.batch_decode(outputs.sequences)[0]
|
120 |
sequence = sequence.replace(task_prompt, "").replace("</s_cord>", "").strip()
|
121 |
|
122 |
try:
|
@@ -124,19 +175,22 @@ def analyze_document(image, model_name, model, processor):
|
|
124 |
except json.JSONDecodeError:
|
125 |
result = {"raw_text": sequence}
|
126 |
|
|
|
|
|
127 |
elif model_name == "LayoutLMv3":
|
128 |
-
#
|
129 |
-
encoded_inputs = processor(
|
130 |
image,
|
131 |
return_tensors="pt",
|
132 |
add_special_tokens=True,
|
133 |
return_offsets_mapping=True
|
134 |
)
|
135 |
|
136 |
-
outputs = model(**encoded_inputs)
|
137 |
predictions = outputs.logits.argmax(-1).squeeze().tolist()
|
138 |
|
139 |
-
|
|
|
140 |
encoded_inputs.input_ids.squeeze().tolist()
|
141 |
)
|
142 |
|
@@ -152,11 +206,19 @@ def analyze_document(image, model_name, model, processor):
|
|
152 |
"confidence_scores": outputs.logits.softmax(-1).max(-1).values.squeeze().tolist()
|
153 |
}
|
154 |
|
155 |
-
|
|
|
|
|
|
|
156 |
|
157 |
except Exception as e:
|
158 |
-
|
159 |
-
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
# Set page config with improved layout
|
162 |
st.set_page_config(
|
@@ -372,6 +434,7 @@ st.markdown("""
|
|
372 |
""")
|
373 |
|
374 |
# Add performance metrics section
|
|
|
375 |
if st.checkbox("Show Performance Metrics"):
|
376 |
st.markdown("""
|
377 |
### Model Performance Metrics
|
@@ -379,8 +442,7 @@ if st.checkbox("Show Performance Metrics"):
|
|
379 |
|-------|---------------------|--------------|-----------|
|
380 |
| Donut | 2-3 seconds | 6-8GB | 85-90% |
|
381 |
| LayoutLMv3 | 3-4 seconds | 12-15GB | 88-93% |
|
382 |
-
|
|
383 |
-
| LLaVA-1.5 | 4-5 seconds | 25-40GB | 90-95% |
|
384 |
|
385 |
*Accuracy varies based on document type and quality
|
386 |
""")
|
@@ -389,7 +451,7 @@ if st.checkbox("Show Performance Metrics"):
|
|
389 |
st.markdown("---")
|
390 |
st.markdown("""
|
391 |
v1.1 - Created with Streamlit
|
392 |
-
\
|
393 |
""")
|
394 |
|
395 |
# Add model selection guidance
|
@@ -398,6 +460,5 @@ if st.checkbox("Show Model Selection Guide"):
|
|
398 |
### How to Choose the Right Model
|
399 |
1. **Donut**: Choose for structured documents with clear layouts
|
400 |
2. **LayoutLMv3**: Best for documents with complex layouts and relationships
|
401 |
-
3. **
|
402 |
-
4. **LLaVA-1.5**: Perfect for complex documents requiring deep understanding
|
403 |
""")
|
|
|
17 |
|
18 |
@st.cache_resource
|
19 |
def load_model(model_name):
|
20 |
+
"""Load the selected model and processor
|
21 |
+
|
22 |
+
Args:
|
23 |
+
model_name (str): Name of the model to load ("Donut", "LayoutLMv3", or "OmniParser")
|
24 |
+
|
25 |
+
Returns:
|
26 |
+
dict: Dictionary containing model components
|
27 |
+
"""
|
28 |
try:
|
29 |
if model_name == "Donut":
|
30 |
processor = DonutProcessor.from_pretrained("naver-clova-ix/donut-base")
|
|
|
34 |
model.config.pad_token_id = processor.tokenizer.pad_token_id
|
35 |
model.config.vocab_size = len(processor.tokenizer)
|
36 |
|
37 |
+
return {'model': model, 'processor': processor}
|
38 |
+
|
39 |
elif model_name == "LayoutLMv3":
|
40 |
processor = LayoutLMv3Processor.from_pretrained("microsoft/layoutlmv3-base")
|
41 |
model = LayoutLMv3ForSequenceClassification.from_pretrained("microsoft/layoutlmv3-base")
|
42 |
|
43 |
+
return {'model': model, 'processor': processor}
|
44 |
+
|
45 |
elif model_name == "OmniParser":
|
46 |
# Load YOLO model for icon detection
|
47 |
+
yolo_model = YOLO("microsoft/OmniParser")
|
48 |
+
|
49 |
+
# Load Florence-2 processor and model for captioning
|
50 |
+
processor = AutoProcessor.from_pretrained(
|
51 |
+
"microsoft/Florence-2-base",
|
52 |
+
trust_remote_code=True
|
53 |
+
)
|
54 |
+
|
55 |
+
# Load the captioning model
|
56 |
+
caption_model = AutoModelForCausalLM.from_pretrained(
|
57 |
+
"microsoft/OmniParser",
|
58 |
trust_remote_code=True
|
59 |
)
|
60 |
|
61 |
return {
|
62 |
'yolo': yolo_model,
|
63 |
'processor': processor,
|
64 |
+
'model': caption_model
|
65 |
}
|
66 |
|
67 |
+
else:
|
68 |
+
raise ValueError(f"Unknown model name: {model_name}")
|
69 |
+
|
70 |
except Exception as e:
|
71 |
st.error(f"Error loading model {model_name}: {str(e)}")
|
72 |
+
return None
|
73 |
|
74 |
+
@spaces.GPU
|
75 |
+
@torch.inference_mode()
|
76 |
+
def analyze_document(image, model_name, models_dict):
|
77 |
+
"""Analyze document using selected model
|
78 |
+
|
79 |
+
Args:
|
80 |
+
image (PIL.Image): Input image to analyze
|
81 |
+
model_name (str): Name of the model to use ("Donut", "LayoutLMv3", or "OmniParser")
|
82 |
+
models_dict (dict): Dictionary containing loaded model components
|
83 |
+
|
84 |
+
Returns:
|
85 |
+
dict: Analysis results including detected elements, text, and/or coordinates
|
86 |
+
"""
|
87 |
try:
|
88 |
+
if models_dict is None:
|
89 |
+
return {"error": "Model failed to load", "type": "model_error"}
|
90 |
+
|
91 |
if model_name == "OmniParser":
|
92 |
+
# Configure detection parameters
|
93 |
+
box_threshold = 0.05 # Confidence threshold for detection
|
94 |
+
iou_threshold = 0.1 # IoU threshold for NMS
|
95 |
+
|
96 |
+
# Save image temporarily for YOLO processing
|
97 |
temp_path = "temp_image.png"
|
98 |
image.save(temp_path)
|
99 |
|
|
|
|
|
|
|
|
|
100 |
# Run YOLO detection
|
101 |
+
yolo_results = models_dict['yolo'](
|
102 |
temp_path,
|
103 |
conf=box_threshold,
|
104 |
iou=iou_threshold
|
105 |
)
|
106 |
|
107 |
+
# Process detections and generate captions
|
108 |
results = []
|
109 |
for det in yolo_results[0].boxes.data:
|
110 |
x1, y1, x2, y2, conf, cls = det
|
111 |
|
112 |
# Get region of interest
|
113 |
+
roi = image.crop((int(x1), int(y1), int(x2), int(y2)))
|
114 |
|
115 |
# Generate caption using the model
|
116 |
+
inputs = models_dict['processor'](
|
117 |
+
images=roi,
|
118 |
+
return_tensors="pt"
|
119 |
+
)
|
120 |
+
|
121 |
+
outputs = models_dict['model'].generate(
|
122 |
+
**inputs,
|
123 |
+
max_length=50,
|
124 |
+
num_beams=4,
|
125 |
+
temperature=0.7
|
126 |
+
)
|
127 |
+
|
128 |
+
caption = models_dict['processor'].decode(outputs[0], skip_special_tokens=True)
|
129 |
|
130 |
results.append({
|
131 |
"bbox": [float(x) for x in [x1, y1, x2, y2]],
|
|
|
134 |
"caption": caption
|
135 |
})
|
136 |
|
137 |
+
# Clean up temporary file
|
138 |
+
if os.path.exists(temp_path):
|
139 |
+
os.remove(temp_path)
|
140 |
+
|
141 |
return {
|
142 |
"detected_elements": len(results),
|
143 |
"elements": results
|
144 |
}
|
145 |
|
146 |
elif model_name == "Donut":
|
147 |
+
# Process image with Donut
|
148 |
+
pixel_values = models_dict['processor'](image, return_tensors="pt").pixel_values
|
149 |
+
|
150 |
task_prompt = "<s_cord>analyze the document and extract information</s_cord>"
|
151 |
+
decoder_input_ids = models_dict['processor'].tokenizer(
|
152 |
+
task_prompt,
|
153 |
+
add_special_tokens=False,
|
154 |
+
return_tensors="pt"
|
155 |
+
).input_ids
|
156 |
|
157 |
+
outputs = models_dict['model'].generate(
|
158 |
pixel_values,
|
159 |
decoder_input_ids=decoder_input_ids,
|
160 |
max_length=512,
|
161 |
early_stopping=True,
|
162 |
+
pad_token_id=models_dict['processor'].tokenizer.pad_token_id,
|
163 |
+
eos_token_id=models_dict['processor'].tokenizer.eos_token_id,
|
164 |
use_cache=True,
|
165 |
num_beams=4,
|
166 |
+
bad_words_ids=[[models_dict['processor'].tokenizer.unk_token_id]],
|
167 |
return_dict_in_generate=True
|
168 |
)
|
169 |
|
170 |
+
sequence = models_dict['processor'].batch_decode(outputs.sequences)[0]
|
171 |
sequence = sequence.replace(task_prompt, "").replace("</s_cord>", "").strip()
|
172 |
|
173 |
try:
|
|
|
175 |
except json.JSONDecodeError:
|
176 |
result = {"raw_text": sequence}
|
177 |
|
178 |
+
return result
|
179 |
+
|
180 |
elif model_name == "LayoutLMv3":
|
181 |
+
# Process image with LayoutLMv3
|
182 |
+
encoded_inputs = models_dict['processor'](
|
183 |
image,
|
184 |
return_tensors="pt",
|
185 |
add_special_tokens=True,
|
186 |
return_offsets_mapping=True
|
187 |
)
|
188 |
|
189 |
+
outputs = models_dict['model'](**encoded_inputs)
|
190 |
predictions = outputs.logits.argmax(-1).squeeze().tolist()
|
191 |
|
192 |
+
# Convert predictions to labels
|
193 |
+
words = models_dict['processor'].tokenizer.convert_ids_to_tokens(
|
194 |
encoded_inputs.input_ids.squeeze().tolist()
|
195 |
)
|
196 |
|
|
|
206 |
"confidence_scores": outputs.logits.softmax(-1).max(-1).values.squeeze().tolist()
|
207 |
}
|
208 |
|
209 |
+
return result
|
210 |
+
|
211 |
+
else:
|
212 |
+
return {"error": f"Unknown model: {model_name}", "type": "model_error"}
|
213 |
|
214 |
except Exception as e:
|
215 |
+
import traceback
|
216 |
+
error_details = traceback.format_exc()
|
217 |
+
return {
|
218 |
+
"error": str(e),
|
219 |
+
"type": "processing_error",
|
220 |
+
"details": error_details
|
221 |
+
}
|
222 |
|
223 |
# Set page config with improved layout
|
224 |
st.set_page_config(
|
|
|
434 |
""")
|
435 |
|
436 |
# Add performance metrics section
|
437 |
+
|
438 |
if st.checkbox("Show Performance Metrics"):
|
439 |
st.markdown("""
|
440 |
### Model Performance Metrics
|
|
|
442 |
|-------|---------------------|--------------|-----------|
|
443 |
| Donut | 2-3 seconds | 6-8GB | 85-90% |
|
444 |
| LayoutLMv3 | 3-4 seconds | 12-15GB | 88-93% |
|
445 |
+
| OmniParser | 2-3 seconds | 8-10GB | 85-90% |
|
|
|
446 |
|
447 |
*Accuracy varies based on document type and quality
|
448 |
""")
|
|
|
451 |
st.markdown("---")
|
452 |
st.markdown("""
|
453 |
v1.1 - Created with Streamlit
|
454 |
+
\nPowered by Hugging Face Spaces 🤗
|
455 |
""")
|
456 |
|
457 |
# Add model selection guidance
|
|
|
460 |
### How to Choose the Right Model
|
461 |
1. **Donut**: Choose for structured documents with clear layouts
|
462 |
2. **LayoutLMv3**: Best for documents with complex layouts and relationships
|
463 |
+
3. **OmniParser**: Best for UI elements and screen parsing
|
|
|
464 |
""")
|