), and others. Include category tags, and filter out irrelevant elements like headers and footers."}]}]
prompt_full = current_processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = current_processor(text=prompt_full, images=[image.convert("RGB")], return_tensors="pt").to(device)
with torch.no_grad():
generated_ids = current_model.generate(**inputs, max_new_tokens=2048, do_sample=False)
generated_ids = generated_ids[:, inputs['input_ids'].shape[1]:]
output_text = current_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
return output_text
elif model_name == "DeepSeek-OCR":
# Move model to the correct device for inference
model_2.to(device)
conversation = [
{"role": "user", "content": ["", image]},
]
input_tensor = tokenizer_2.apply_chat_template(conversation, return_tensors="pt")
with torch.no_grad():
output_tensor = model_2.run(input_tensor.to(device))
# This model returns plain text, so we wrap it in basic HTML for consistency
ocr_text = output_tensor[0]
html_output = "".join(f"{line}
" for line in ocr_text.split('\n'))
return html_output
else:
raise ValueError(f"Unknown model choice: {model_name}")
def convert_file_to_images(file_path: str, dpi: int = 200) -> List[Image.Image]:
images = []
file_ext = Path(file_path).suffix.lower()
if file_ext in image_suffixes:
images.append(Image.open(file_path).convert("RGB"))
return images
if file_ext not in pdf_suffixes:
raise ValueError(f"Unsupported file type: {file_ext}")
try:
pdf_document = fitz.open(file_path)
zoom = dpi / 72.0
mat = fitz.Matrix(zoom, zoom)
for page_num in range(len(pdf_document)):
page = pdf_document.load_page(page_num)
pix = page.get_pixmap(matrix=mat)
img_data = pix.tobytes("png")
images.append(Image.open(BytesIO(img_data)).convert("RGB"))
pdf_document.close()
except Exception as e:
logger.error(f"Failed to convert PDF using PyMuPDF: {e}")
raise
return images
def get_initial_state() -> Dict[str, Any]:
return {"pages": [], "total_pages": 0, "current_page_index": 0, "page_results": []}
def load_and_preview_file(file_path: Optional[str]) -> Tuple[Optional[Image.Image], str, Dict[str, Any]]:
state = get_initial_state()
if not file_path:
return None, 'No file loaded
', state
try:
pages = convert_file_to_images(file_path)
if not pages:
return None, 'Could not load file
', state
state["pages"] = pages
state["total_pages"] = len(pages)
page_info_html = f'Page 1 / {state["total_pages"]}
'
return pages[0], page_info_html, state
except Exception as e:
logger.error(f"Failed to load and preview file: {e}")
return None, 'Failed to load preview
', state
async def process_all_pages(state: Dict[str, Any], model_choice: str, progress=gr.Progress(track_tqdm=True)):
if not state or not state["pages"]:
error_msg = "Please upload a file first.
"
return error_msg, "", "", None, "Error: No file to process", state
logger.info(f'Processing {state["total_pages"]} pages with model: {model_choice}')
start_time = time.time()
try:
page_results = []
for i, page_img in progress.tqdm(enumerate(state["pages"]), desc="Processing Pages"):
html_result = parse_page(page_img, model_choice)
page_results.append({'raw_html': html_result})
state["page_results"] = page_results
full_html_content = "\n\n".join([f'\n{res["raw_html"]}' for i, res in enumerate(page_results)])
full_markdown = html2text.html2text(full_html_content)
with tempfile.NamedTemporaryFile(mode='w', suffix='.md', delete=False, encoding='utf-8') as f:
f.write(full_markdown)
md_path = f.name
parsing_time = time.time() - start_time
cost_time_str = f'Total processing time: {parsing_time:.2f}s'
current_page_results = get_page_outputs(state)
return *current_page_results, md_path, cost_time_str, state
except Exception as e:
logger.error(f"Parsing failed: {e}", exc_info=True)
error_html = f"An error occurred during processing:
{str(e)}
"
return error_html, "", "", None, f"Error: {str(e)}", state
def navigate_page(direction: str, state: Dict[str, Any]):
if not state or not state["pages"]:
return None, 'No file loaded
', *get_page_outputs(state), state
current_index = state["current_page_index"]
total_pages = state["total_pages"]
if direction == "prev":
new_index = max(0, current_index - 1)
elif direction == "next":
new_index = min(total_pages - 1, current_index + 1)
else:
new_index = current_index
state["current_page_index"] = new_index
image_preview = state["pages"][new_index]
page_info_html = f'Page {new_index + 1} / {total_pages}
'
page_outputs = get_page_outputs(state)
return image_preview, page_info_html, *page_outputs, state
def get_page_outputs(state: Dict[str, Any]) -> Tuple[str, str, str]:
if not state or not state.get("page_results"):
return "Process the document to see results.
", "", ""
index = state["current_page_index"]
if index >= len(state["page_results"]):
return "Result not available for this page.
", "", ""
result = state["page_results"][index]
raw_html = result['raw_html']
md_source = html2text.html2text(raw_html)
md_render = markdown.markdown(md_source, extensions=['fenced_code', 'tables'])
return md_render, md_source, raw_html
def clear_all():
return None, None, "Results will be displayed here after processing.
", "", "", None, "", 'No file loaded
', get_initial_state()
@click.command()
def main():
css = """
.main-container { max-width: 1400px; margin: 0 auto; }
.header-text { text-align: center; margin-bottom: 20px; }
.page-info { text-align: center; padding: 8px 16px; font-weight: bold; margin: 10px 0; }
"""
with gr.Blocks(theme=steel_blue_theme, css=css, title="Logics-Parsing Demo") as demo:
app_state = gr.State(value=get_initial_state())
gr.HTML("""
""")
with gr.Row(elem_classes=["main-container"]):
with gr.Column(scale=1):
model_choice = gr.Dropdown(choices=["Logics-Parsing", "DeepSeek-OCR"], label="Select Model", value="Logics-Parsing")
file_input = gr.File(label="Upload PDF or Image", file_types=[".pdf", ".jpg", ".jpeg", ".png"], type="filepath")
process_btn = gr.Button("πProcess Document", variant="primary", size="lg")
clear_btn = gr.Button("ποΈ Clear All", variant="secondary")
image_preview = gr.Image(label="Preview", type="pil", interactive=False, height=320)
with gr.Row():
prev_page_btn = gr.Button("β Previous")
page_info = gr.HTML('No file loaded
')
next_page_btn = gr.Button("Next βΆ")
example_root = "examples"
if os.path.exists(example_root) and os.path.isdir(example_root):
example_files = [os.path.join(example_root, f) for f in os.listdir(example_root) if f.endswith(tuple(pdf_suffixes + image_suffixes))]
if example_files:
gr.Examples(examples=example_files, inputs=file_input, label="Examples")
with gr.Accordion("Download & Details", open=False):
output_file = gr.File(label='Download Markdown Result', interactive=False)
cost_time = gr.Textbox(label='Time Cost', interactive=False)
with gr.Column(scale=2):
with gr.Tabs():
with gr.Tab("Markdown Source"):
md_source_output = gr.Code(language="markdown", label="Markdown Source")
with gr.Tab("Rendered Markdown"):
md_render_output = gr.Markdown(label='Markdown Rendering')
with gr.Tab("Generated HTML"):
raw_html_output = gr.Code(language="html", label="Generated HTML")
file_input.change(fn=load_and_preview_file, inputs=file_input, outputs=[image_preview, page_info, app_state], show_progress="full")
process_btn.click(fn=process_all_pages, inputs=[app_state, model_choice], outputs=[md_render_output, md_source_output, raw_html_output, output_file, cost_time, app_state], show_progress="full")
prev_page_btn.click(fn=lambda s: navigate_page("prev", s), inputs=app_state, outputs=[image_preview, page_info, md_render_output, md_source_output, raw_html_output, app_state])
next_page_btn.click(fn=lambda s: navigate_page("next", s), inputs=app_state, outputs=[image_preview, page_info, md_render_output, md_source_output, raw_html_output, app_state])
clear_btn.click(fn=clear_all, outputs=[file_input, image_preview, md_render_output, md_source_output, raw_html_output, output_file, cost_time, page_info, app_state])
demo.queue().launch(debug=True, show_error=True)
if __name__ == '__main__':
if not os.path.exists("examples"):
os.makedirs("examples")
logger.info("Created 'examples' directory. Please add some sample PDF/image files there.")
main()