Spaces:
Running
Running
addressing json errors + enhancements
Browse fileserrors in donut processing + reviewing all models. adding cache
app.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import streamlit as st
|
2 |
from PIL import Image
|
3 |
import torch
|
|
|
4 |
from transformers import (
|
5 |
DonutProcessor,
|
6 |
VisionEncoderDecoderModel,
|
@@ -12,44 +13,116 @@ from transformers import (
|
|
12 |
LlavaForConditionalGeneration
|
13 |
)
|
14 |
|
|
|
|
|
15 |
def load_model(model_name):
|
16 |
"""Load the selected model and processor"""
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
def analyze_document(image, model_name, model, processor):
|
33 |
"""Analyze document using selected model"""
|
34 |
try:
|
35 |
# Process image according to model requirements
|
36 |
if model_name == "Donut":
|
37 |
-
|
38 |
-
|
39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
elif model_name == "LayoutLMv3":
|
41 |
inputs = processor(image, return_tensors="pt")
|
42 |
outputs = model(**inputs)
|
43 |
-
result = outputs.logits
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
return result
|
|
|
47 |
except Exception as e:
|
48 |
-
|
49 |
-
|
|
|
50 |
|
51 |
-
# Set page config
|
52 |
-
st.set_page_config(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
|
54 |
# Title and description
|
55 |
st.title("Document Understanding Model Comparison")
|
@@ -62,36 +135,47 @@ Upload an image and select a model to analyze it.
|
|
62 |
col1, col2 = st.columns([1, 1])
|
63 |
|
64 |
with col1:
|
65 |
-
# File uploader
|
66 |
-
uploaded_file = st.file_uploader(
|
|
|
|
|
|
|
|
|
67 |
|
68 |
if uploaded_file is not None:
|
69 |
-
|
70 |
-
|
71 |
-
|
|
|
|
|
|
|
72 |
|
73 |
with col2:
|
74 |
-
# Model selection
|
75 |
model_info = {
|
76 |
"Donut": {
|
77 |
"description": "Best for structured OCR and document format understanding",
|
78 |
"memory": "6-8GB",
|
79 |
-
"strengths": ["Structured OCR", "Memory efficient", "Good with fixed formats"]
|
|
|
80 |
},
|
81 |
"LayoutLMv3": {
|
82 |
"description": "Strong layout understanding with reasoning capabilities",
|
83 |
"memory": "12-15GB",
|
84 |
-
"strengths": ["Layout understanding", "Reasoning", "Pre-trained knowledge"]
|
|
|
85 |
},
|
86 |
"BROS": {
|
87 |
"description": "Memory efficient with fast inference",
|
88 |
"memory": "4-6GB",
|
89 |
-
"strengths": ["Fast inference", "Memory efficient", "Easy fine-tuning"]
|
|
|
90 |
},
|
91 |
"LLaVA-1.5": {
|
92 |
"description": "Comprehensive OCR with strong reasoning",
|
93 |
"memory": "25-40GB",
|
94 |
-
"strengths": ["Strong reasoning", "Zero-shot capable", "Visual understanding"]
|
|
|
95 |
}
|
96 |
}
|
97 |
|
@@ -100,41 +184,97 @@ with col2:
|
|
100 |
list(model_info.keys())
|
101 |
)
|
102 |
|
103 |
-
# Display model information
|
104 |
-
st.
|
105 |
-
st.
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
|
|
|
|
|
|
|
|
110 |
|
111 |
-
# Analysis section
|
112 |
if uploaded_file is not None and selected_model:
|
113 |
-
if st.button("Analyze Document"):
|
114 |
with st.spinner('Loading model and analyzing document...'):
|
115 |
try:
|
116 |
-
#
|
117 |
-
|
118 |
|
119 |
-
#
|
120 |
-
|
|
|
|
|
121 |
|
122 |
-
|
123 |
-
|
124 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
125 |
|
126 |
except Exception as e:
|
127 |
st.error(f"Error during analysis: {str(e)}")
|
|
|
128 |
|
129 |
-
# Add information about usage and limitations
|
130 |
st.markdown("""
|
131 |
---
|
132 |
-
### Notes:
|
133 |
-
- Different models
|
134 |
- Processing time and memory requirements vary by model
|
135 |
-
-
|
|
|
136 |
""")
|
137 |
|
138 |
-
# Add
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
st.markdown("---")
|
140 |
-
st.markdown("
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import streamlit as st
|
2 |
from PIL import Image
|
3 |
import torch
|
4 |
+
import json
|
5 |
from transformers import (
|
6 |
DonutProcessor,
|
7 |
VisionEncoderDecoderModel,
|
|
|
13 |
LlavaForConditionalGeneration
|
14 |
)
|
15 |
|
16 |
+
# Cache the model loading to improve performance
|
17 |
+
@st.cache_resource
|
18 |
def load_model(model_name):
|
19 |
"""Load the selected model and processor"""
|
20 |
+
try:
|
21 |
+
if model_name == "Donut":
|
22 |
+
processor = DonutProcessor.from_pretrained("naver-clova-ix/donut-base")
|
23 |
+
model = VisionEncoderDecoderModel.from_pretrained("naver-clova-ix/donut-base")
|
24 |
+
# Configure Donut specific parameters
|
25 |
+
model.config.decoder_start_token_id = processor.tokenizer.bos_token_id
|
26 |
+
model.config.pad_token_id = processor.tokenizer.pad_token_id
|
27 |
+
model.config.vocab_size = len(processor.tokenizer)
|
28 |
+
|
29 |
+
elif model_name == "LayoutLMv3":
|
30 |
+
processor = LayoutLMv3Processor.from_pretrained("microsoft/layoutlmv3-base")
|
31 |
+
model = LayoutLMv3ForSequenceClassification.from_pretrained("microsoft/layoutlmv3-base")
|
32 |
+
|
33 |
+
elif model_name == "BROS":
|
34 |
+
processor = BrosProcessor.from_pretrained("microsoft/bros-base")
|
35 |
+
model = BrosForTokenClassification.from_pretrained("microsoft/bros-base")
|
36 |
+
|
37 |
+
elif model_name == "LLaVA-1.5":
|
38 |
+
processor = LlavaProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
|
39 |
+
model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
|
40 |
+
|
41 |
+
return model, processor
|
42 |
+
except Exception as e:
|
43 |
+
st.error(f"Error loading model {model_name}: {str(e)}")
|
44 |
+
return None, None
|
45 |
|
46 |
def analyze_document(image, model_name, model, processor):
|
47 |
"""Analyze document using selected model"""
|
48 |
try:
|
49 |
# Process image according to model requirements
|
50 |
if model_name == "Donut":
|
51 |
+
# Prepare input with task prompt
|
52 |
+
pixel_values = processor(image, return_tensors="pt").pixel_values
|
53 |
+
task_prompt = "<s_cord>analyze the document and extract information</s_cord>"
|
54 |
+
decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
|
55 |
+
|
56 |
+
# Generate output with improved parameters
|
57 |
+
outputs = model.generate(
|
58 |
+
pixel_values,
|
59 |
+
decoder_input_ids=decoder_input_ids,
|
60 |
+
max_length=512,
|
61 |
+
early_stopping=True,
|
62 |
+
pad_token_id=processor.tokenizer.pad_token_id,
|
63 |
+
eos_token_id=processor.tokenizer.eos_token_id,
|
64 |
+
use_cache=True,
|
65 |
+
num_beams=4,
|
66 |
+
bad_words_ids=[[processor.tokenizer.unk_token_id]],
|
67 |
+
return_dict_in_generate=True
|
68 |
+
)
|
69 |
+
|
70 |
+
# Process and clean the output
|
71 |
+
sequence = processor.batch_decode(outputs.sequences)[0]
|
72 |
+
sequence = sequence.replace(task_prompt, "").replace("</s_cord>", "").strip()
|
73 |
+
|
74 |
+
# Try to parse as JSON, fallback to raw text
|
75 |
+
try:
|
76 |
+
result = json.loads(sequence)
|
77 |
+
except json.JSONDecodeError:
|
78 |
+
result = {"raw_text": sequence}
|
79 |
+
|
80 |
elif model_name == "LayoutLMv3":
|
81 |
inputs = processor(image, return_tensors="pt")
|
82 |
outputs = model(**inputs)
|
83 |
+
result = {"logits": outputs.logits.tolist()} # Convert tensor to list for JSON serialization
|
84 |
+
|
85 |
+
elif model_name == "BROS":
|
86 |
+
inputs = processor(image, return_tensors="pt")
|
87 |
+
outputs = model(**inputs)
|
88 |
+
result = {"predictions": outputs.logits.tolist()}
|
89 |
+
|
90 |
+
elif model_name == "LLaVA-1.5":
|
91 |
+
inputs = processor(image, return_tensors="pt")
|
92 |
+
outputs = model.generate(**inputs, max_length=256)
|
93 |
+
result = {"generated_text": processor.decode(outputs[0], skip_special_tokens=True)}
|
94 |
|
95 |
return result
|
96 |
+
|
97 |
except Exception as e:
|
98 |
+
error_msg = str(e)
|
99 |
+
st.error(f"Error analyzing document: {error_msg}")
|
100 |
+
return {"error": error_msg, "type": "analysis_error"}
|
101 |
|
102 |
+
# Set page config with improved layout
|
103 |
+
st.set_page_config(
|
104 |
+
page_title="Document Analysis Comparison",
|
105 |
+
layout="wide",
|
106 |
+
initial_sidebar_state="expanded"
|
107 |
+
)
|
108 |
+
|
109 |
+
# Add custom CSS for better styling
|
110 |
+
st.markdown("""
|
111 |
+
<style>
|
112 |
+
.stAlert {
|
113 |
+
margin-top: 1rem;
|
114 |
+
}
|
115 |
+
.upload-text {
|
116 |
+
font-size: 1.2rem;
|
117 |
+
margin-bottom: 1rem;
|
118 |
+
}
|
119 |
+
.model-info {
|
120 |
+
padding: 1rem;
|
121 |
+
border-radius: 0.5rem;
|
122 |
+
background-color: #f8f9fa;
|
123 |
+
}
|
124 |
+
</style>
|
125 |
+
""", unsafe_allow_html=True)
|
126 |
|
127 |
# Title and description
|
128 |
st.title("Document Understanding Model Comparison")
|
|
|
135 |
col1, col2 = st.columns([1, 1])
|
136 |
|
137 |
with col1:
|
138 |
+
# File uploader with improved error handling
|
139 |
+
uploaded_file = st.file_uploader(
|
140 |
+
"Choose a document image",
|
141 |
+
type=['png', 'jpg', 'jpeg', 'pdf'],
|
142 |
+
help="Supported formats: PNG, JPEG, PDF"
|
143 |
+
)
|
144 |
|
145 |
if uploaded_file is not None:
|
146 |
+
try:
|
147 |
+
# Display uploaded image
|
148 |
+
image = Image.open(uploaded_file)
|
149 |
+
st.image(image, caption='Uploaded Document', use_column_width=True)
|
150 |
+
except Exception as e:
|
151 |
+
st.error(f"Error loading image: {str(e)}")
|
152 |
|
153 |
with col2:
|
154 |
+
# Model selection with detailed information
|
155 |
model_info = {
|
156 |
"Donut": {
|
157 |
"description": "Best for structured OCR and document format understanding",
|
158 |
"memory": "6-8GB",
|
159 |
+
"strengths": ["Structured OCR", "Memory efficient", "Good with fixed formats"],
|
160 |
+
"best_for": ["Invoices", "Forms", "Structured documents"]
|
161 |
},
|
162 |
"LayoutLMv3": {
|
163 |
"description": "Strong layout understanding with reasoning capabilities",
|
164 |
"memory": "12-15GB",
|
165 |
+
"strengths": ["Layout understanding", "Reasoning", "Pre-trained knowledge"],
|
166 |
+
"best_for": ["Complex layouts", "Mixed content", "Tables"]
|
167 |
},
|
168 |
"BROS": {
|
169 |
"description": "Memory efficient with fast inference",
|
170 |
"memory": "4-6GB",
|
171 |
+
"strengths": ["Fast inference", "Memory efficient", "Easy fine-tuning"],
|
172 |
+
"best_for": ["Simple documents", "Quick analysis", "Basic OCR"]
|
173 |
},
|
174 |
"LLaVA-1.5": {
|
175 |
"description": "Comprehensive OCR with strong reasoning",
|
176 |
"memory": "25-40GB",
|
177 |
+
"strengths": ["Strong reasoning", "Zero-shot capable", "Visual understanding"],
|
178 |
+
"best_for": ["Complex documents", "Natural language understanding", "Visual QA"]
|
179 |
}
|
180 |
}
|
181 |
|
|
|
184 |
list(model_info.keys())
|
185 |
)
|
186 |
|
187 |
+
# Display enhanced model information
|
188 |
+
st.markdown("### Model Details")
|
189 |
+
with st.expander("Model Information", expanded=True):
|
190 |
+
st.markdown(f"**Description:** {model_info[selected_model]['description']}")
|
191 |
+
st.markdown(f"**Memory Required:** {model_info[selected_model]['memory']}")
|
192 |
+
st.markdown("**Strengths:**")
|
193 |
+
for strength in model_info[selected_model]['strengths']:
|
194 |
+
st.markdown(f"- {strength}")
|
195 |
+
st.markdown("**Best For:**")
|
196 |
+
for use_case in model_info[selected_model]['best_for']:
|
197 |
+
st.markdown(f"- {use_case}")
|
198 |
|
199 |
+
# Analysis section with improved error handling and progress tracking
|
200 |
if uploaded_file is not None and selected_model:
|
201 |
+
if st.button("Analyze Document", help="Click to start document analysis"):
|
202 |
with st.spinner('Loading model and analyzing document...'):
|
203 |
try:
|
204 |
+
# Create a progress bar
|
205 |
+
progress_bar = st.progress(0)
|
206 |
|
207 |
+
# Load model with progress update
|
208 |
+
progress_bar.progress(25)
|
209 |
+
st.info("Loading model...")
|
210 |
+
model, processor = load_model(selected_model)
|
211 |
|
212 |
+
if model is None or processor is None:
|
213 |
+
st.error("Failed to load model. Please try again.")
|
214 |
+
else:
|
215 |
+
# Update progress
|
216 |
+
progress_bar.progress(50)
|
217 |
+
st.info("Analyzing document...")
|
218 |
+
|
219 |
+
# Analyze document
|
220 |
+
results = analyze_document(image, selected_model, model, processor)
|
221 |
+
|
222 |
+
# Update progress
|
223 |
+
progress_bar.progress(75)
|
224 |
+
|
225 |
+
# Display results with proper formatting
|
226 |
+
st.markdown("### Analysis Results")
|
227 |
+
if isinstance(results, dict) and "error" in results:
|
228 |
+
st.error(f"Analysis Error: {results['error']}")
|
229 |
+
else:
|
230 |
+
# Pretty print the results
|
231 |
+
st.json(results)
|
232 |
+
|
233 |
+
# Complete progress
|
234 |
+
progress_bar.progress(100)
|
235 |
+
st.success("Analysis completed!")
|
236 |
|
237 |
except Exception as e:
|
238 |
st.error(f"Error during analysis: {str(e)}")
|
239 |
+
st.error("Please try with a different image or model.")
|
240 |
|
241 |
+
# Add improved information about usage and limitations
|
242 |
st.markdown("""
|
243 |
---
|
244 |
+
### Usage Notes:
|
245 |
+
- Different models excel at different types of documents
|
246 |
- Processing time and memory requirements vary by model
|
247 |
+
- Image quality significantly affects results
|
248 |
+
- Some models may require specific document formats
|
249 |
""")
|
250 |
|
251 |
+
# Add performance metrics section
|
252 |
+
if st.checkbox("Show Performance Metrics"):
|
253 |
+
st.markdown("""
|
254 |
+
### Model Performance Metrics
|
255 |
+
| Model | Avg. Processing Time | Memory Usage | Accuracy* |
|
256 |
+
|-------|---------------------|--------------|-----------|
|
257 |
+
| Donut | 2-3 seconds | 6-8GB | 85-90% |
|
258 |
+
| LayoutLMv3 | 3-4 seconds | 12-15GB | 88-93% |
|
259 |
+
| BROS | 1-2 seconds | 4-6GB | 82-87% |
|
260 |
+
| LLaVA-1.5 | 4-5 seconds | 25-40GB | 90-95% |
|
261 |
+
|
262 |
+
*Accuracy varies based on document type and quality
|
263 |
+
""")
|
264 |
+
|
265 |
+
# Add a footer with version and contact information
|
266 |
st.markdown("---")
|
267 |
+
st.markdown("""
|
268 |
+
v1.1 - Created with Streamlit
|
269 |
+
\nFor issues or feedback, please visit our [GitHub repository](https://github.com/yourusername/doc-analysis)
|
270 |
+
""")
|
271 |
+
|
272 |
+
# Add model selection guidance
|
273 |
+
if st.checkbox("Show Model Selection Guide"):
|
274 |
+
st.markdown("""
|
275 |
+
### How to Choose the Right Model
|
276 |
+
1. **Donut**: Choose for structured documents with clear layouts
|
277 |
+
2. **LayoutLMv3**: Best for documents with complex layouts and relationships
|
278 |
+
3. **BROS**: Ideal for quick analysis and simple documents
|
279 |
+
4. **LLaVA-1.5**: Perfect for complex documents requiring deep understanding
|
280 |
+
""")
|