Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -6,8 +6,9 @@ import traceback
|
|
6 |
from io import BytesIO
|
7 |
from typing import Any, Dict, List, Optional, Tuple
|
8 |
import re
|
|
|
|
|
9 |
|
10 |
-
import fitz
|
11 |
import gradio as gr
|
12 |
import requests
|
13 |
import torch
|
@@ -17,15 +18,15 @@ from transformers import (
|
|
17 |
AutoProcessor,
|
18 |
TextIteratorStreamer,
|
19 |
)
|
20 |
-
from qwen_vl_utils import process_vision_info
|
21 |
|
22 |
# Constants
|
23 |
MIN_PIXELS = 3136
|
24 |
MAX_PIXELS = 11289600
|
25 |
IMAGE_FACTOR = 28
|
|
|
26 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
27 |
|
28 |
-
#
|
29 |
prompt = """Please output the layout information from the PDF image, including each layout element's bbox, its category, and the corresponding text content within the bbox.
|
30 |
|
31 |
1. Bbox format: [x1, y1, x2, y2]
|
@@ -45,190 +46,74 @@ prompt = """Please output the layout information from the PDF image, including e
|
|
45 |
5. Final Output: The entire output must be a single JSON object.
|
46 |
"""
|
47 |
|
48 |
-
# Load
|
49 |
MODEL_ID_M = "prithivMLmods/Camel-Doc-OCR-062825"
|
50 |
processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
|
51 |
model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
52 |
-
MODEL_ID_M,
|
53 |
-
trust_remote_code=True,
|
54 |
-
torch_dtype=torch.float16
|
55 |
).to(device).eval()
|
56 |
|
57 |
MODEL_ID_T = "prithivMLmods/Megalodon-OCR-Sync-0713"
|
58 |
processor_t = AutoProcessor.from_pretrained(MODEL_ID_T, trust_remote_code=True)
|
59 |
model_t = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
60 |
-
MODEL_ID_T,
|
61 |
-
trust_remote_code=True,
|
62 |
-
torch_dtype=torch.float16
|
63 |
).to(device).eval()
|
64 |
|
65 |
MODEL_ID_C = "nanonets/Nanonets-OCR-s"
|
66 |
processor_c = AutoProcessor.from_pretrained(MODEL_ID_C, trust_remote_code=True)
|
67 |
model_c = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
68 |
-
MODEL_ID_C,
|
69 |
-
trust_remote_code=True,
|
70 |
-
torch_dtype=torch.float16
|
71 |
).to(device).eval()
|
72 |
|
73 |
MODEL_ID_G = "echo840/MonkeyOCR"
|
74 |
SUBFOLDER = "Recognition"
|
75 |
processor_g = AutoProcessor.from_pretrained(
|
76 |
-
MODEL_ID_G,
|
77 |
-
trust_remote_code=True,
|
78 |
-
subfolder=SUBFOLDER
|
79 |
)
|
80 |
model_g = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
81 |
-
MODEL_ID_G,
|
82 |
-
trust_remote_code=True,
|
83 |
-
subfolder=SUBFOLDER,
|
84 |
-
torch_dtype=torch.float16
|
85 |
).to(device).eval()
|
86 |
|
87 |
-
# Utility functions
|
88 |
-
def round_by_factor(number: int, factor: int) -> int:
|
89 |
-
return round(number / factor) * factor
|
90 |
-
|
91 |
-
def smart_resize(
|
92 |
-
height: int,
|
93 |
-
width: int,
|
94 |
-
factor: int = 28,
|
95 |
-
min_pixels: int = 3136,
|
96 |
-
max_pixels: int = 11289600,
|
97 |
-
):
|
98 |
-
if max(height, width) / min(height, width) > 200:
|
99 |
-
raise ValueError(f"Aspect ratio too extreme: {max(height, width) / min(height, width)}")
|
100 |
-
h_bar = max(factor, round_by_factor(height, factor))
|
101 |
-
w_bar = max(factor, round_by_factor(width, factor))
|
102 |
-
if h_bar * w_bar > max_pixels:
|
103 |
-
beta = math.sqrt((height * width) / max_pixels)
|
104 |
-
h_bar = round_by_factor(height / beta, factor)
|
105 |
-
w_bar = round_by_factor(width / beta, factor)
|
106 |
-
elif h_bar * w_bar < min_pixels:
|
107 |
-
beta = math.sqrt(min_pixels / (height * width))
|
108 |
-
h_bar = round_by_factor(height * beta, factor)
|
109 |
-
w_bar = round_by_factor(width * beta, factor)
|
110 |
-
return h_bar, w_bar
|
111 |
-
|
112 |
-
def fetch_image(image_input, min_pixels: int = None, max_pixels: int = None):
|
113 |
-
if isinstance(image_input, str):
|
114 |
-
if image_input.startswith(("http://", "https://")):
|
115 |
-
response = requests.get(image_input)
|
116 |
-
image = Image.open(BytesIO(response.content)).convert('RGB')
|
117 |
-
else:
|
118 |
-
image = Image.open(image_input).convert('RGB')
|
119 |
-
elif isinstance(image_input, Image.Image):
|
120 |
-
image = image_input.convert('RGB')
|
121 |
-
else:
|
122 |
-
raise ValueError(f"Invalid image input type: {type(image_input)}")
|
123 |
-
if min_pixels or max_pixels:
|
124 |
-
min_pixels = min_pixels or MIN_PIXELS
|
125 |
-
max_pixels = max_pixels or MAX_PIXELS
|
126 |
-
height, width = smart_resize(image.height, image.width, factor=IMAGE_FACTOR, min_pixels=min_pixels, max_pixels=max_pixels)
|
127 |
-
image = image.resize((width, height), Image.LANCZOS)
|
128 |
-
return image
|
129 |
-
|
130 |
-
def load_images_from_pdf(pdf_path: str) -> List[Image.Image]:
|
131 |
-
images = []
|
132 |
-
try:
|
133 |
-
pdf_document = fitz.open(pdf_path)
|
134 |
-
for page_num in range(len(pdf_document)):
|
135 |
-
page = pdf_document.load_page(page_num)
|
136 |
-
mat = fitz.Matrix(2.0, 2.0)
|
137 |
-
pix = page.get_pixmap(matrix=mat)
|
138 |
-
img_data = pix.tobytes("ppm")
|
139 |
-
image = Image.open(BytesIO(img_data)).convert('RGB')
|
140 |
-
images.append(image)
|
141 |
-
pdf_document.close()
|
142 |
-
except Exception as e:
|
143 |
-
print(f"Error loading PDF: {e}")
|
144 |
-
return images
|
145 |
-
|
146 |
-
def draw_layout_on_image(image: Image.Image, layout_data: List[Dict]) -> Image.Image:
|
147 |
-
img_copy = image.copy()
|
148 |
-
draw = ImageDraw.Draw(img_copy)
|
149 |
-
colors = {
|
150 |
-
'Caption': '#FF6B6B', 'Footnote': '#4ECDC4', 'Formula': '#45B7D1',
|
151 |
-
'List-item': '#96CEB4', 'Page-footer': '#FFEAA7', 'Page-header': '#DDA0DD',
|
152 |
-
'Picture': '#FFD93D', 'Section-header': '#6C5CE7', 'Table': '#FD79A8',
|
153 |
-
'Text': '#74B9FF', 'Title': '#E17055'
|
154 |
-
}
|
155 |
-
try:
|
156 |
-
font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 12)
|
157 |
-
except Exception:
|
158 |
-
font = ImageFont.load_default()
|
159 |
-
for item in layout_data:
|
160 |
-
if 'bbox' in item and 'category' in item:
|
161 |
-
bbox = item['bbox']
|
162 |
-
category = item['category']
|
163 |
-
color = colors.get(category, '#000000')
|
164 |
-
draw.rectangle(bbox, outline=color, width=2)
|
165 |
-
label = category
|
166 |
-
label_bbox = draw.textbbox((0, 0), label, font=font)
|
167 |
-
label_width = label_bbox[2] - label_bbox[0]
|
168 |
-
label_height = label_bbox[3] - label_bbox[1]
|
169 |
-
label_x = bbox[0]
|
170 |
-
label_y = max(0, bbox[1] - label_height - 2)
|
171 |
-
draw.rectangle([label_x, label_y, label_x + label_width + 4, label_y + label_height + 2], fill=color)
|
172 |
-
draw.text((label_x + 2, label_y + 1), label, fill='white', font=font)
|
173 |
-
return img_copy
|
174 |
|
|
|
175 |
def is_arabic_text(text: str) -> bool:
|
|
|
176 |
if not text:
|
177 |
return False
|
178 |
-
|
179 |
-
paragraph_pattern = r'^(?!#{1,6}\s|!\[|```|\||\s*[-*+]\s|\s*\d+\.\s)(.+)$'
|
180 |
-
content_text = []
|
181 |
-
for line in text.split('\n'):
|
182 |
-
line = line.strip()
|
183 |
-
if not line:
|
184 |
-
continue
|
185 |
-
header_match = re.match(header_pattern, line, re.MULTILINE)
|
186 |
-
if header_match:
|
187 |
-
content_text.append(header_match.group(1))
|
188 |
-
continue
|
189 |
-
if re.match(paragraph_pattern, line, re.MULTILINE):
|
190 |
-
content_text.append(line)
|
191 |
-
if not content_text:
|
192 |
-
return False
|
193 |
-
combined_text = ' '.join(content_text)
|
194 |
arabic_chars = 0
|
195 |
total_chars = 0
|
196 |
-
for char in
|
197 |
if char.isalpha():
|
198 |
total_chars += 1
|
199 |
-
if
|
200 |
arabic_chars += 1
|
201 |
return total_chars > 0 and (arabic_chars / total_chars) > 0.5
|
202 |
|
203 |
def layoutjson2md(image: Image.Image, layout_data: List[Dict], text_key: str = 'text') -> str:
|
|
|
204 |
import base64
|
205 |
from io import BytesIO
|
206 |
markdown_lines = []
|
207 |
try:
|
|
|
208 |
sorted_items = sorted(layout_data, key=lambda x: (x.get('bbox', [0, 0, 0, 0])[1], x.get('bbox', [0, 0, 0, 0])[0]))
|
209 |
for item in sorted_items:
|
210 |
category = item.get('category', '')
|
211 |
text = item.get(text_key, '')
|
212 |
bbox = item.get('bbox', [])
|
|
|
213 |
if category == 'Picture':
|
214 |
if bbox and len(bbox) == 4:
|
215 |
try:
|
216 |
-
x1, y1, x2, y2 = bbox
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
cropped_img.save(buffer, format='PNG')
|
223 |
-
img_data = base64.b64encode(buffer.getvalue()).decode()
|
224 |
-
markdown_lines.append(f"\n")
|
225 |
-
else:
|
226 |
-
markdown_lines.append("\n")
|
227 |
except Exception as e:
|
228 |
-
|
229 |
-
markdown_lines.append("\n")
|
230 |
-
else:
|
231 |
-
markdown_lines.append("\n")
|
232 |
elif not text:
|
233 |
continue
|
234 |
elif category == 'Title':
|
@@ -239,311 +124,173 @@ def layoutjson2md(image: Image.Image, layout_data: List[Dict], text_key: str = '
|
|
239 |
markdown_lines.append(f"{text}\n")
|
240 |
elif category == 'List-item':
|
241 |
markdown_lines.append(f"- {text}\n")
|
242 |
-
elif category == 'Table':
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
-
markdown_lines.append(f"**Table:** {text}\n")
|
247 |
-
elif category == 'Formula':
|
248 |
-
if text.strip().startswith('$') or '\\' in text:
|
249 |
-
markdown_lines.append(f"$$\n{text}\n$$\n")
|
250 |
-
else:
|
251 |
-
markdown_lines.append(f"**Formula:** {text}\n")
|
252 |
elif category == 'Caption':
|
253 |
markdown_lines.append(f"*{text}*\n")
|
254 |
elif category == 'Footnote':
|
255 |
-
|
256 |
-
elif category in ['Page-header', 'Page-footer']:
|
257 |
-
continue
|
258 |
-
else:
|
259 |
markdown_lines.append(f"{text}\n")
|
260 |
-
markdown_lines.append("")
|
261 |
except Exception as e:
|
262 |
print(f"Error converting to markdown: {e}")
|
263 |
-
return str(layout_data)
|
264 |
return "\n".join(markdown_lines)
|
265 |
|
266 |
-
# PDF handling state
|
267 |
-
pdf_cache = {
|
268 |
-
"images": [],
|
269 |
-
"current_page": 0,
|
270 |
-
"total_pages": 0,
|
271 |
-
"file_type": None,
|
272 |
-
"is_parsed": False,
|
273 |
-
"results": []
|
274 |
-
}
|
275 |
|
276 |
@spaces.GPU
|
277 |
-
def
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
elif model_name == "Nanonets-OCR-s":
|
286 |
-
processor = processor_c
|
287 |
-
model = model_c
|
288 |
-
elif model_name == "MonkeyOCR-Recognition":
|
289 |
-
processor = processor_g
|
290 |
-
model = model_g
|
291 |
-
else:
|
292 |
-
raise ValueError(f"Invalid model selected: {model_name}")
|
293 |
-
|
294 |
-
messages = [
|
295 |
-
{
|
296 |
-
"role": "user",
|
297 |
-
"content": [
|
298 |
-
{"type": "image", "image": image},
|
299 |
-
{"type": "text", "text": prompt}
|
300 |
-
]
|
301 |
-
}
|
302 |
-
]
|
303 |
-
|
304 |
-
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
305 |
-
inputs = processor(text=[text], images=[image], return_tensors="pt").to(device)
|
306 |
-
|
307 |
-
with torch.no_grad():
|
308 |
-
generated_ids = model.generate(
|
309 |
-
**inputs,
|
310 |
-
max_new_tokens=max_new_tokens,
|
311 |
-
do_sample=False,
|
312 |
-
temperature=0.1
|
313 |
-
)
|
314 |
-
|
315 |
-
generated_ids_trimmed = [
|
316 |
-
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
317 |
-
]
|
318 |
-
output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
319 |
-
return output_text[0] if output_text else ""
|
320 |
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
|
340 |
-
|
341 |
-
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
|
361 |
-
|
362 |
-
|
363 |
-
except Exception as e:
|
364 |
-
print(f"Error processing image: {e}")
|
365 |
-
traceback.print_exc()
|
366 |
-
return {
|
367 |
-
'original_image': image,
|
368 |
-
'raw_output': f"Error processing image: {str(e)}",
|
369 |
-
'processed_image': image,
|
370 |
-
'layout_result': None,
|
371 |
-
'markdown_content': f"Error processing image: {str(e)}"
|
372 |
-
}
|
373 |
-
|
374 |
-
def load_file_for_preview(file_path: str) -> Tuple[Optional[Image.Image], str]:
|
375 |
-
global pdf_cache
|
376 |
-
if not file_path or not os.path.exists(file_path):
|
377 |
-
return None, "No file selected"
|
378 |
-
file_ext = os.path.splitext(file_path)[1].lower()
|
379 |
try:
|
380 |
-
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
|
392 |
-
return images[0], f"Page 1 / {len(images)}"
|
393 |
-
elif file_ext in ['.jpg', '.jpeg', '.png', '.bmp', '.tiff']:
|
394 |
-
image = Image.open(file_path).convert('RGB')
|
395 |
-
pdf_cache.update({
|
396 |
-
"images": [image],
|
397 |
-
"current_page": 0,
|
398 |
-
"total_pages": 1,
|
399 |
-
"file_type": "image",
|
400 |
-
"is_parsed": False,
|
401 |
-
"results": []
|
402 |
-
})
|
403 |
-
return image, "Page 1 / 1"
|
404 |
-
else:
|
405 |
-
return None, f"Unsupported file format: {file_ext}"
|
406 |
except Exception as e:
|
407 |
-
|
408 |
-
|
409 |
-
|
410 |
-
def turn_page(direction: str) -> Tuple[Optional[Image.Image], str, Any, Optional[Image.Image], Optional[Dict]]:
|
411 |
-
global pdf_cache
|
412 |
-
if not pdf_cache["images"]:
|
413 |
-
return None, '<div class="page-info">No file loaded</div>', "No results yet", None, None
|
414 |
-
if direction == "prev":
|
415 |
-
pdf_cache["current_page"] = max(0, pdf_cache["current_page"] - 1)
|
416 |
-
elif direction == "next":
|
417 |
-
pdf_cache["current_page"] = min(pdf_cache["total_pages"] - 1, pdf_cache["current_page"] + 1)
|
418 |
-
index = pdf_cache["current_page"]
|
419 |
-
current_image_preview = pdf_cache["images"][index]
|
420 |
-
page_info_html = f'<div class="page-info">Page {index + 1} / {pdf_cache["total_pages"]}</div>'
|
421 |
-
markdown_content = "Page not processed yet"
|
422 |
-
processed_img = None
|
423 |
-
layout_json = None
|
424 |
-
if pdf_cache["is_parsed"] and index < len(pdf_cache["results"]) and pdf_cache["results"][index]:
|
425 |
-
result = pdf_cache["results"][index]
|
426 |
-
markdown_content = result.get('markdown_content') or result.get('raw_output', 'No content available')
|
427 |
-
processed_img = result.get('processed_image', None)
|
428 |
-
layout_json = result.get('layout_result', None)
|
429 |
-
if is_arabic_text(markdown_content):
|
430 |
-
markdown_update = gr.update(value=markdown_content, rtl=True)
|
431 |
-
else:
|
432 |
-
markdown_update = markdown_content
|
433 |
-
return current_image_preview, page_info_html, markdown_update, processed_img, layout_json
|
434 |
|
435 |
def create_gradio_interface():
|
|
|
436 |
css = """
|
437 |
.main-container { max-width: 1400px; margin: 0 auto; }
|
438 |
.header-text { text-align: center; color: #2c3e50; margin-bottom: 20px; }
|
439 |
.process-button {
|
440 |
-
border: none !important;
|
441 |
-
color:
|
442 |
-
|
443 |
-
background-color: blue !important;}
|
444 |
.process-button:hover {
|
445 |
-
background-color: darkblue !important;
|
446 |
-
|
447 |
-
|
448 |
-
.info-box { border: 1px solid #dee2e6; border-radius: 8px; padding: 15px; margin: 10px 0; }
|
449 |
-
.page-info { text-align: center; padding: 8px 16px; border-radius: 20px; font-weight: bold; margin: 10px 0; }
|
450 |
-
.model-status { padding: 10px; border-radius: 8px; margin: 10px 0; text-align: center; font-weight: bold; }
|
451 |
-
.status-ready { background: #d1edff; color: #0c5460; border: 1px solid #b8daff; }
|
452 |
"""
|
453 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css) as demo:
|
454 |
gr.HTML("""
|
455 |
<div class="title" style="text-align: center">
|
456 |
<h1>Dot<span style="color: red;">β</span><strong></strong>OCR Comparator</h1>
|
457 |
<p style="font-size: 1.1em; color: #6b7280; margin-bottom: 0.6em;">
|
458 |
-
Advanced vision-language model for image
|
459 |
</p>
|
460 |
</div>
|
461 |
""")
|
|
|
|
|
|
|
|
|
462 |
with gr.Row():
|
|
|
463 |
with gr.Column(scale=1):
|
464 |
model_choice = gr.Radio(
|
465 |
choices=["Camel-Doc-OCR-062825", "MonkeyOCR-Recognition", "Nanonets-OCR-s", "Megalodon-OCR-Sync-0713"],
|
466 |
label="Select Model",
|
467 |
value="Camel-Doc-OCR-062825"
|
468 |
)
|
469 |
-
file_input = gr.
|
470 |
-
label="Upload Image
|
471 |
-
|
472 |
-
|
473 |
)
|
474 |
-
image_preview = gr.Image(label="Preview", type="pil", interactive=False, height=300)
|
475 |
-
with gr.Row():
|
476 |
-
prev_page_btn = gr.Button("β Previous", size="md")
|
477 |
-
page_info = gr.HTML('<div class="page-info">No file loaded</div>')
|
478 |
-
next_page_btn = gr.Button("Next βΆ", size="md")
|
479 |
with gr.Accordion("Advanced Settings", open=False):
|
480 |
max_new_tokens = gr.Slider(minimum=1000, maximum=32000, value=24000, step=1000, label="Max New Tokens")
|
481 |
-
|
482 |
-
max_pixels = gr.Number(value=MAX_PIXELS, label="Max Pixels")
|
483 |
process_btn = gr.Button("π Process Document", variant="primary", elem_classes=["process-button"], size="lg")
|
484 |
clear_btn = gr.Button("ποΈ Clear All", variant="secondary")
|
|
|
|
|
485 |
with gr.Column(scale=2):
|
486 |
with gr.Tabs():
|
487 |
-
with gr.Tab("πΌοΈ Processed Image"):
|
488 |
-
processed_image = gr.Image(label="Image with Layout Detection", type="pil", interactive=False, height=500)
|
489 |
with gr.Tab("π Extracted Content"):
|
490 |
-
|
|
|
|
|
|
|
491 |
with gr.Tab("π Layout JSON"):
|
492 |
-
json_output = gr.JSON(label="Layout Analysis Results", value=None)
|
493 |
-
|
494 |
-
|
495 |
-
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
if pdf_cache["file_type"] == "pdf":
|
500 |
-
all_results = []
|
501 |
-
all_markdown = []
|
502 |
-
for i, img in enumerate(pdf_cache["images"]):
|
503 |
-
result = process_image(model_name, img, min_pixels=int(min_pix) if min_pix else None, max_pixels=int(max_pix) if max_pix else None)
|
504 |
-
all_results.append(result)
|
505 |
-
if result.get('markdown_content'):
|
506 |
-
all_markdown.append(f"## Page {i+1}\n\n{result['markdown_content']}")
|
507 |
-
pdf_cache["results"] = all_results
|
508 |
-
pdf_cache["is_parsed"] = True
|
509 |
-
first_result = all_results[0]
|
510 |
-
combined_markdown = "\n\n---\n\n".join(all_markdown)
|
511 |
-
if is_arabic_text(combined_markdown):
|
512 |
-
markdown_update = gr.update(value=combined_markdown, rtl=True)
|
513 |
-
else:
|
514 |
-
markdown_update = combined_markdown
|
515 |
-
return first_result['processed_image'], markdown_update, first_result['layout_result']
|
516 |
-
else:
|
517 |
-
result = process_image(model_name, pdf_cache["images"][0], min_pixels=int(min_pix) if min_pix else None, max_pixels=int(max_pix) if max_pix else None)
|
518 |
-
pdf_cache["results"] = [result]
|
519 |
-
pdf_cache["is_parsed"] = True
|
520 |
-
content = result['markdown_content'] or "No content extracted"
|
521 |
-
if is_arabic_text(content):
|
522 |
-
markdown_update = gr.update(value=content, rtl=True)
|
523 |
-
else:
|
524 |
-
markdown_update = content
|
525 |
-
return result['processed_image'], markdown_update, result['layout_result']
|
526 |
-
except Exception as e:
|
527 |
-
error_msg = f"Error processing document: {str(e)}"
|
528 |
-
print(error_msg)
|
529 |
-
traceback.print_exc()
|
530 |
-
return None, error_msg, None
|
531 |
-
def handle_file_upload(file_path):
|
532 |
-
if not file_path:
|
533 |
-
return None, "No file loaded"
|
534 |
-
image, page_info = load_file_for_preview(file_path)
|
535 |
-
return image, page_info
|
536 |
def clear_all():
|
537 |
-
|
538 |
-
|
539 |
-
|
540 |
-
file_input.
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
545 |
return demo
|
546 |
|
547 |
if __name__ == "__main__":
|
548 |
demo = create_gradio_interface()
|
549 |
-
demo.queue(
|
|
|
6 |
from io import BytesIO
|
7 |
from typing import Any, Dict, List, Optional, Tuple
|
8 |
import re
|
9 |
+
from threading import Thread
|
10 |
+
import time
|
11 |
|
|
|
12 |
import gradio as gr
|
13 |
import requests
|
14 |
import torch
|
|
|
18 |
AutoProcessor,
|
19 |
TextIteratorStreamer,
|
20 |
)
|
|
|
21 |
|
22 |
# Constants
|
23 |
MIN_PIXELS = 3136
|
24 |
MAX_PIXELS = 11289600
|
25 |
IMAGE_FACTOR = 28
|
26 |
+
MAX_INPUT_TOKEN_LENGTH = 4096
|
27 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
28 |
|
29 |
+
# Prompt for Layout Analysis
|
30 |
prompt = """Please output the layout information from the PDF image, including each layout element's bbox, its category, and the corresponding text content within the bbox.
|
31 |
|
32 |
1. Bbox format: [x1, y1, x2, y2]
|
|
|
46 |
5. Final Output: The entire output must be a single JSON object.
|
47 |
"""
|
48 |
|
49 |
+
# Load Models
|
50 |
MODEL_ID_M = "prithivMLmods/Camel-Doc-OCR-062825"
|
51 |
processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
|
52 |
model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
53 |
+
MODEL_ID_M, trust_remote_code=True, torch_dtype=torch.float16
|
|
|
|
|
54 |
).to(device).eval()
|
55 |
|
56 |
MODEL_ID_T = "prithivMLmods/Megalodon-OCR-Sync-0713"
|
57 |
processor_t = AutoProcessor.from_pretrained(MODEL_ID_T, trust_remote_code=True)
|
58 |
model_t = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
59 |
+
MODEL_ID_T, trust_remote_code=True, torch_dtype=torch.float16
|
|
|
|
|
60 |
).to(device).eval()
|
61 |
|
62 |
MODEL_ID_C = "nanonets/Nanonets-OCR-s"
|
63 |
processor_c = AutoProcessor.from_pretrained(MODEL_ID_C, trust_remote_code=True)
|
64 |
model_c = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
65 |
+
MODEL_ID_C, trust_remote_code=True, torch_dtype=torch.float16
|
|
|
|
|
66 |
).to(device).eval()
|
67 |
|
68 |
MODEL_ID_G = "echo840/MonkeyOCR"
|
69 |
SUBFOLDER = "Recognition"
|
70 |
processor_g = AutoProcessor.from_pretrained(
|
71 |
+
MODEL_ID_G, trust_remote_code=True, subfolder=SUBFOLDER
|
|
|
|
|
72 |
)
|
73 |
model_g = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
74 |
+
MODEL_ID_G, trust_remote_code=True, subfolder=SUBFOLDER, torch_dtype=torch.float16
|
|
|
|
|
|
|
75 |
).to(device).eval()
|
76 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
|
78 |
+
# Utility functions
|
79 |
def is_arabic_text(text: str) -> bool:
|
80 |
+
"""Check if text contains mostly Arabic characters."""
|
81 |
if not text:
|
82 |
return False
|
83 |
+
# Simplified check for Arabic characters in the given text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
arabic_chars = 0
|
85 |
total_chars = 0
|
86 |
+
for char in text:
|
87 |
if char.isalpha():
|
88 |
total_chars += 1
|
89 |
+
if '\u0600' <= char <= '\u06FF':
|
90 |
arabic_chars += 1
|
91 |
return total_chars > 0 and (arabic_chars / total_chars) > 0.5
|
92 |
|
93 |
def layoutjson2md(image: Image.Image, layout_data: List[Dict], text_key: str = 'text') -> str:
|
94 |
+
"""Convert layout JSON to markdown format."""
|
95 |
import base64
|
96 |
from io import BytesIO
|
97 |
markdown_lines = []
|
98 |
try:
|
99 |
+
# Sort items by reading order (top to bottom, left to right)
|
100 |
sorted_items = sorted(layout_data, key=lambda x: (x.get('bbox', [0, 0, 0, 0])[1], x.get('bbox', [0, 0, 0, 0])[0]))
|
101 |
for item in sorted_items:
|
102 |
category = item.get('category', '')
|
103 |
text = item.get(text_key, '')
|
104 |
bbox = item.get('bbox', [])
|
105 |
+
|
106 |
if category == 'Picture':
|
107 |
if bbox and len(bbox) == 4:
|
108 |
try:
|
109 |
+
x1, y1, x2, y2 = [int(coord) for coord in bbox]
|
110 |
+
cropped_img = image.crop((x1, y1, x2, y2))
|
111 |
+
buffer = BytesIO()
|
112 |
+
cropped_img.save(buffer, format='PNG')
|
113 |
+
img_data = base64.b64encode(buffer.getvalue()).decode()
|
114 |
+
markdown_lines.append(f"\n")
|
|
|
|
|
|
|
|
|
|
|
115 |
except Exception as e:
|
116 |
+
markdown_lines.append("\n")
|
|
|
|
|
|
|
117 |
elif not text:
|
118 |
continue
|
119 |
elif category == 'Title':
|
|
|
124 |
markdown_lines.append(f"{text}\n")
|
125 |
elif category == 'List-item':
|
126 |
markdown_lines.append(f"- {text}\n")
|
127 |
+
elif category == 'Table' and text.strip().startswith('<'):
|
128 |
+
markdown_lines.append(f"{text}\n")
|
129 |
+
elif category == 'Formula' and (text.strip().startswith('$') or '\\' in text):
|
130 |
+
markdown_lines.append(f"$$\n{text}\n$$\n")
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
elif category == 'Caption':
|
132 |
markdown_lines.append(f"*{text}*\n")
|
133 |
elif category == 'Footnote':
|
134 |
+
markdown_lines.append(f"^{text}^\n")
|
135 |
+
elif category not in ['Page-header', 'Page-footer']:
|
|
|
|
|
136 |
markdown_lines.append(f"{text}\n")
|
|
|
137 |
except Exception as e:
|
138 |
print(f"Error converting to markdown: {e}")
|
139 |
+
return f"### Error converting to Markdown\n\n```\n{str(layout_data)}\n```"
|
140 |
return "\n".join(markdown_lines)
|
141 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
|
143 |
@spaces.GPU
|
144 |
+
def generate_and_process(model_name: str, image: Image.Image, max_new_tokens: int):
|
145 |
+
"""
|
146 |
+
Generates a response using streaming, then processes the final output.
|
147 |
+
Yields updates for the raw stream, final markdown, and JSON output.
|
148 |
+
"""
|
149 |
+
if image is None:
|
150 |
+
yield "Please upload an image.", "Please upload an image.", None
|
151 |
+
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
+
# 1. Select Model and Processor
|
154 |
+
if model_name == "Camel-Doc-OCR-062825":
|
155 |
+
processor, model = processor_m, model_m
|
156 |
+
elif model_name == "Megalodon-OCR-Sync-0713":
|
157 |
+
processor, model = processor_t, model_t
|
158 |
+
elif model_name == "Nanonets-OCR-s":
|
159 |
+
processor, model = processor_c, model_c
|
160 |
+
elif model_name == "MonkeyOCR-Recognition":
|
161 |
+
processor, model = processor_g, model_g
|
162 |
+
else:
|
163 |
+
yield "Invalid model selected.", "Invalid model selected.", None
|
164 |
+
return
|
165 |
+
|
166 |
+
# 2. Prepare inputs for the model
|
167 |
+
messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}]
|
168 |
+
prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
169 |
+
inputs = processor(
|
170 |
+
text=[prompt_full],
|
171 |
+
images=[image],
|
172 |
+
return_tensors="pt",
|
173 |
+
padding=True,
|
174 |
+
truncation=True,
|
175 |
+
max_length=MAX_INPUT_TOKEN_LENGTH
|
176 |
+
).to(device)
|
177 |
+
|
178 |
+
# 3. Stream the generation
|
179 |
+
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
180 |
+
generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
|
181 |
+
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
182 |
+
thread.start()
|
183 |
+
|
184 |
+
buffer = ""
|
185 |
+
# Initial placeholder yield
|
186 |
+
yield buffer, "β³ Generating response...", None
|
187 |
+
|
188 |
+
for new_text in streamer:
|
189 |
+
buffer += new_text
|
190 |
+
buffer = buffer.replace("<|im_end|>", "")
|
191 |
+
time.sleep(0.01) # Small delay for smoother streaming
|
192 |
+
yield buffer, "β³ Generating response...", None
|
193 |
+
|
194 |
+
# 4. Process the final buffer content
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
195 |
try:
|
196 |
+
json_match = re.search(r'```json\s*([\s\S]+?)\s*```', buffer)
|
197 |
+
json_str = json_match.group(1) if json_match else buffer
|
198 |
+
layout_data = json.loads(json_str)
|
199 |
+
|
200 |
+
markdown_content = layoutjson2md(image, layout_data)
|
201 |
+
|
202 |
+
# Final yield with all processed content
|
203 |
+
yield buffer, markdown_content, layout_data
|
204 |
+
|
205 |
+
except json.JSONDecodeError:
|
206 |
+
error_msg = "β Failed to parse JSON from model output."
|
207 |
+
yield buffer, error_msg, {"error": "JSONDecodeError", "raw_output": buffer}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
208 |
except Exception as e:
|
209 |
+
error_msg = f"β An error occurred during post-processing: {e}"
|
210 |
+
yield buffer, error_msg, {"error": str(e), "raw_output": buffer}
|
211 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
212 |
|
213 |
def create_gradio_interface():
|
214 |
+
"""Create the Gradio interface."""
|
215 |
css = """
|
216 |
.main-container { max-width: 1400px; margin: 0 auto; }
|
217 |
.header-text { text-align: center; color: #2c3e50; margin-bottom: 20px; }
|
218 |
.process-button {
|
219 |
+
border: none !important; color: white !important; font-weight: bold !important;
|
220 |
+
background-color: blue !important;
|
221 |
+
}
|
|
|
222 |
.process-button:hover {
|
223 |
+
background-color: darkblue !important; transform: translateY(-2px) !important;
|
224 |
+
box-shadow: 0 4px 8px rgba(0,0,0,0.2) !important;
|
225 |
+
}
|
|
|
|
|
|
|
|
|
226 |
"""
|
227 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css) as demo:
|
228 |
gr.HTML("""
|
229 |
<div class="title" style="text-align: center">
|
230 |
<h1>Dot<span style="color: red;">β</span><strong></strong>OCR Comparator</h1>
|
231 |
<p style="font-size: 1.1em; color: #6b7280; margin-bottom: 0.6em;">
|
232 |
+
Advanced vision-language model for image to markdown document processing
|
233 |
</p>
|
234 |
</div>
|
235 |
""")
|
236 |
+
|
237 |
+
# Keep track of the uploaded image
|
238 |
+
image_state = gr.State(None)
|
239 |
+
|
240 |
with gr.Row():
|
241 |
+
# Left column - Input and controls
|
242 |
with gr.Column(scale=1):
|
243 |
model_choice = gr.Radio(
|
244 |
choices=["Camel-Doc-OCR-062825", "MonkeyOCR-Recognition", "Nanonets-OCR-s", "Megalodon-OCR-Sync-0713"],
|
245 |
label="Select Model",
|
246 |
value="Camel-Doc-OCR-062825"
|
247 |
)
|
248 |
+
file_input = gr.Image(
|
249 |
+
label="Upload Image",
|
250 |
+
type="pil",
|
251 |
+
sources=['upload']
|
252 |
)
|
|
|
|
|
|
|
|
|
|
|
253 |
with gr.Accordion("Advanced Settings", open=False):
|
254 |
max_new_tokens = gr.Slider(minimum=1000, maximum=32000, value=24000, step=1000, label="Max New Tokens")
|
255 |
+
|
|
|
256 |
process_btn = gr.Button("π Process Document", variant="primary", elem_classes=["process-button"], size="lg")
|
257 |
clear_btn = gr.Button("ποΈ Clear All", variant="secondary")
|
258 |
+
|
259 |
+
# Right column - Results
|
260 |
with gr.Column(scale=2):
|
261 |
with gr.Tabs():
|
|
|
|
|
262 |
with gr.Tab("π Extracted Content"):
|
263 |
+
output_stream = gr.Textbox(label="Raw Output Stream", interactive=False, lines=10, show_copy_button=True)
|
264 |
+
with gr.Accordion("(Formatted Result)", open=True):
|
265 |
+
markdown_output = gr.Markdown(label="Formatted Result (Result.md)")
|
266 |
+
|
267 |
with gr.Tab("π Layout JSON"):
|
268 |
+
json_output = gr.JSON(label="Layout Analysis Results (JSON)", value=None)
|
269 |
+
|
270 |
+
# Event Handlers
|
271 |
+
def handle_file_upload(image):
|
272 |
+
"""Store the uploaded image in the state."""
|
273 |
+
return image
|
274 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
275 |
def clear_all():
|
276 |
+
"""Clear all data and reset the interface."""
|
277 |
+
return None, None, "Click 'Process Document' to see extracted content...", None, None
|
278 |
+
|
279 |
+
file_input.upload(handle_file_upload, inputs=[file_input], outputs=[image_state])
|
280 |
+
|
281 |
+
process_btn.click(
|
282 |
+
generate_and_process,
|
283 |
+
inputs=[model_choice, image_state, max_new_tokens],
|
284 |
+
outputs=[output_stream, markdown_output, json_output]
|
285 |
+
)
|
286 |
+
|
287 |
+
clear_btn.click(
|
288 |
+
clear_all,
|
289 |
+
outputs=[file_input, image_state, markdown_output, json_output, output_stream]
|
290 |
+
)
|
291 |
+
|
292 |
return demo
|
293 |
|
294 |
if __name__ == "__main__":
|
295 |
demo = create_gradio_interface()
|
296 |
+
demo.queue().launch(server_name="0.0.0.0", server_port=7860, share=True, show_error=True)
|