Spaces:
Running
Running
donut str error & omniparser path error
Browse files
app.py
CHANGED
@@ -45,17 +45,17 @@ def load_model(model_name):
|
|
45 |
|
46 |
elif model_name == "OmniParser":
|
47 |
# Load YOLO model for icon detection
|
48 |
-
yolo_model = YOLO("microsoft/OmniParser")
|
49 |
|
50 |
# Load Florence-2 processor and model for captioning
|
51 |
processor = AutoProcessor.from_pretrained(
|
52 |
-
"microsoft/
|
53 |
trust_remote_code=True
|
54 |
)
|
55 |
|
56 |
# Load the captioning model
|
57 |
caption_model = AutoModelForCausalLM.from_pretrained(
|
58 |
-
"microsoft/OmniParser",
|
59 |
trust_remote_code=True
|
60 |
)
|
61 |
|
@@ -75,16 +75,7 @@ def load_model(model_name):
|
|
75 |
@spaces.GPU
|
76 |
@torch.inference_mode()
|
77 |
def analyze_document(image, model_name, models_dict):
|
78 |
-
"""Analyze document using selected model
|
79 |
-
|
80 |
-
Args:
|
81 |
-
image (PIL.Image): Input image to analyze
|
82 |
-
model_name (str): Name of the model to use ("Donut", "LayoutLMv3", or "OmniParser")
|
83 |
-
models_dict (dict): Dictionary containing loaded model components
|
84 |
-
|
85 |
-
Returns:
|
86 |
-
dict: Analysis results including detected elements, text, and/or coordinates
|
87 |
-
"""
|
88 |
try:
|
89 |
if models_dict is None:
|
90 |
return {"error": "Model failed to load", "type": "model_error"}
|
@@ -98,77 +89,82 @@ def analyze_document(image, model_name, models_dict):
|
|
98 |
temp_path = "temp_image.png"
|
99 |
image.save(temp_path)
|
100 |
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
# Process detections and generate captions
|
109 |
-
results = []
|
110 |
-
for det in yolo_results[0].boxes.data:
|
111 |
-
x1, y1, x2, y2, conf, cls = det
|
112 |
-
|
113 |
-
# Get region of interest
|
114 |
-
roi = image.crop((int(x1), int(y1), int(x2), int(y2)))
|
115 |
-
|
116 |
-
# Generate caption using the model
|
117 |
-
inputs = models_dict['processor'](
|
118 |
-
images=roi,
|
119 |
-
return_tensors="pt"
|
120 |
-
)
|
121 |
-
|
122 |
-
outputs = models_dict['model'].generate(
|
123 |
-
**inputs,
|
124 |
-
max_length=50,
|
125 |
-
num_beams=4,
|
126 |
-
temperature=0.7
|
127 |
)
|
128 |
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
"caption": caption
|
136 |
-
})
|
137 |
-
|
138 |
-
# Clean up temporary file
|
139 |
-
if os.path.exists(temp_path):
|
140 |
-
os.remove(temp_path)
|
141 |
-
|
142 |
-
return {
|
143 |
-
"detected_elements": len(results),
|
144 |
-
"elements": results
|
145 |
-
}
|
146 |
|
147 |
elif model_name == "Donut":
|
|
|
|
|
|
|
148 |
# Process image with Donut
|
149 |
-
pixel_values =
|
150 |
|
151 |
task_prompt = "<s_cord>analyze the document and extract information</s_cord>"
|
152 |
-
decoder_input_ids =
|
153 |
task_prompt,
|
154 |
add_special_tokens=False,
|
155 |
return_tensors="pt"
|
156 |
).input_ids
|
157 |
|
158 |
-
outputs =
|
159 |
pixel_values,
|
160 |
decoder_input_ids=decoder_input_ids,
|
161 |
max_length=512,
|
162 |
early_stopping=True,
|
163 |
-
pad_token_id=
|
164 |
-
eos_token_id=
|
165 |
use_cache=True,
|
166 |
num_beams=4,
|
167 |
-
bad_words_ids=[[
|
168 |
return_dict_in_generate=True
|
169 |
)
|
170 |
|
171 |
-
sequence =
|
172 |
sequence = sequence.replace(task_prompt, "").replace("</s_cord>", "").strip()
|
173 |
|
174 |
try:
|
@@ -179,19 +175,22 @@ def analyze_document(image, model_name, models_dict):
|
|
179 |
return result
|
180 |
|
181 |
elif model_name == "LayoutLMv3":
|
|
|
|
|
|
|
182 |
# Process image with LayoutLMv3
|
183 |
-
encoded_inputs =
|
184 |
image,
|
185 |
return_tensors="pt",
|
186 |
add_special_tokens=True,
|
187 |
return_offsets_mapping=True
|
188 |
)
|
189 |
|
190 |
-
outputs =
|
191 |
predictions = outputs.logits.argmax(-1).squeeze().tolist()
|
192 |
|
193 |
# Convert predictions to labels
|
194 |
-
words =
|
195 |
encoded_inputs.input_ids.squeeze().tolist()
|
196 |
)
|
197 |
|
@@ -215,6 +214,7 @@ def analyze_document(image, model_name, models_dict):
|
|
215 |
except Exception as e:
|
216 |
import traceback
|
217 |
error_details = traceback.format_exc()
|
|
|
218 |
return {
|
219 |
"error": str(e),
|
220 |
"type": "processing_error",
|
|
|
45 |
|
46 |
elif model_name == "OmniParser":
|
47 |
# Load YOLO model for icon detection
|
48 |
+
yolo_model = YOLO("microsoft/OmniParser-icon-detection")
|
49 |
|
50 |
# Load Florence-2 processor and model for captioning
|
51 |
processor = AutoProcessor.from_pretrained(
|
52 |
+
"microsoft/OmniParser-caption",
|
53 |
trust_remote_code=True
|
54 |
)
|
55 |
|
56 |
# Load the captioning model
|
57 |
caption_model = AutoModelForCausalLM.from_pretrained(
|
58 |
+
"microsoft/OmniParser-caption",
|
59 |
trust_remote_code=True
|
60 |
)
|
61 |
|
|
|
75 |
@spaces.GPU
|
76 |
@torch.inference_mode()
|
77 |
def analyze_document(image, model_name, models_dict):
|
78 |
+
"""Analyze document using selected model"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
try:
|
80 |
if models_dict is None:
|
81 |
return {"error": "Model failed to load", "type": "model_error"}
|
|
|
89 |
temp_path = "temp_image.png"
|
90 |
image.save(temp_path)
|
91 |
|
92 |
+
try:
|
93 |
+
# Run YOLO detection
|
94 |
+
yolo_results = models_dict['yolo'](
|
95 |
+
temp_path,
|
96 |
+
conf=box_threshold,
|
97 |
+
iou=iou_threshold
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
)
|
99 |
|
100 |
+
# Process detections and generate captions
|
101 |
+
results = []
|
102 |
+
for det in yolo_results[0].boxes.data:
|
103 |
+
x1, y1, x2, y2, conf, cls = det
|
104 |
+
|
105 |
+
# Get region of interest
|
106 |
+
roi = image.crop((int(x1), int(y1), int(x2), int(y2)))
|
107 |
+
|
108 |
+
# Generate caption using the model
|
109 |
+
inputs = models_dict['processor'](
|
110 |
+
images=roi,
|
111 |
+
return_tensors="pt"
|
112 |
+
)
|
113 |
+
|
114 |
+
outputs = models_dict['model'].generate(
|
115 |
+
**inputs,
|
116 |
+
max_length=50,
|
117 |
+
num_beams=4,
|
118 |
+
temperature=0.7
|
119 |
+
)
|
120 |
+
|
121 |
+
caption = models_dict['processor'].decode(outputs[0], skip_special_tokens=True)
|
122 |
+
|
123 |
+
results.append({
|
124 |
+
"bbox": [float(x) for x in [x1, y1, x2, y2]],
|
125 |
+
"confidence": float(conf),
|
126 |
+
"class": int(cls),
|
127 |
+
"caption": caption
|
128 |
+
})
|
129 |
+
|
130 |
+
return {
|
131 |
+
"detected_elements": len(results),
|
132 |
+
"elements": results
|
133 |
+
}
|
134 |
|
135 |
+
finally:
|
136 |
+
# Clean up temporary file
|
137 |
+
if os.path.exists(temp_path):
|
138 |
+
os.remove(temp_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
|
140 |
elif model_name == "Donut":
|
141 |
+
model = models_dict['model']
|
142 |
+
processor = models_dict['processor']
|
143 |
+
|
144 |
# Process image with Donut
|
145 |
+
pixel_values = processor(image, return_tensors="pt").pixel_values
|
146 |
|
147 |
task_prompt = "<s_cord>analyze the document and extract information</s_cord>"
|
148 |
+
decoder_input_ids = processor.tokenizer(
|
149 |
task_prompt,
|
150 |
add_special_tokens=False,
|
151 |
return_tensors="pt"
|
152 |
).input_ids
|
153 |
|
154 |
+
outputs = model.generate(
|
155 |
pixel_values,
|
156 |
decoder_input_ids=decoder_input_ids,
|
157 |
max_length=512,
|
158 |
early_stopping=True,
|
159 |
+
pad_token_id=processor.tokenizer.pad_token_id,
|
160 |
+
eos_token_id=processor.tokenizer.eos_token_id,
|
161 |
use_cache=True,
|
162 |
num_beams=4,
|
163 |
+
bad_words_ids=[[processor.tokenizer.unk_token_id]],
|
164 |
return_dict_in_generate=True
|
165 |
)
|
166 |
|
167 |
+
sequence = processor.batch_decode(outputs.sequences)[0]
|
168 |
sequence = sequence.replace(task_prompt, "").replace("</s_cord>", "").strip()
|
169 |
|
170 |
try:
|
|
|
175 |
return result
|
176 |
|
177 |
elif model_name == "LayoutLMv3":
|
178 |
+
model = models_dict['model']
|
179 |
+
processor = models_dict['processor']
|
180 |
+
|
181 |
# Process image with LayoutLMv3
|
182 |
+
encoded_inputs = processor(
|
183 |
image,
|
184 |
return_tensors="pt",
|
185 |
add_special_tokens=True,
|
186 |
return_offsets_mapping=True
|
187 |
)
|
188 |
|
189 |
+
outputs = model(**encoded_inputs)
|
190 |
predictions = outputs.logits.argmax(-1).squeeze().tolist()
|
191 |
|
192 |
# Convert predictions to labels
|
193 |
+
words = processor.tokenizer.convert_ids_to_tokens(
|
194 |
encoded_inputs.input_ids.squeeze().tolist()
|
195 |
)
|
196 |
|
|
|
214 |
except Exception as e:
|
215 |
import traceback
|
216 |
error_details = traceback.format_exc()
|
217 |
+
logger.error(f"Analysis error: {str(e)}\n{error_details}")
|
218 |
return {
|
219 |
"error": str(e),
|
220 |
"type": "processing_error",
|