dots-ocr / app.py
mrbubbaz's picture
Update app.py
640db2e verified
Raw History Blame Contribute Delete
8.9 kB
"""
dots.ocr 1.5 β€” Hugging Face Spaces (GPU T4/A10G)
Deploy: crea un nuovo Space su huggingface.co/new-space
- SDK: Gradio
- Hardware: T4 GPU (free) o A10G
- Carica questo file come app.py + requirements.txt
"""
import os
import json
import torch
import gradio as gr
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor
from qwen_vl_utils import process_vision_info
# ── Configurazione ──────────────────────────────────────────────
MODEL_ID = "rednote-hilab/dots.ocr-1.5"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.bfloat16 if DEVICE == "cuda" else torch.float32
print(f"πŸ–₯️ Device: {DEVICE} | dtype: {DTYPE}")
# ── Patch: mock flash_attn se non installato ────────────────────
import importlib
if not importlib.util.find_spec("flash_attn"):
import types
flash_attn_mock = types.ModuleType("flash_attn")
flash_attn_mock.flash_attn_varlen_func = None
import sys
sys.modules["flash_attn"] = flash_attn_mock
print("⚠️ flash_attn non disponibile, mock installato (verrà usato sdpa)")
# ── Caricamento modello ─────────────────────────────────────────
print(f"πŸ“¦ Caricamento modello {MODEL_ID}...")
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
dtype=DTYPE,
device_map="auto",
trust_remote_code=True,
attn_implementation="sdpa", # usa SDPA invece di flash_attention_2
)
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
print("βœ… Modello caricato!")
# ── Prompt ──────────────────────────────────────────────────────
PROMPTS = {
"πŸ“„ Parsing completo (layout + testo)": """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.
1. Bbox format: [x1, y1, x2, y2]
2. Layout Categories: The possible categories are ['Caption', 'Footnote', 'Formula', 'List-item', 'Page-footer', 'Page-header', 'Picture', 'Section-header', 'Table', 'Text', 'Title'].
3. Text Extraction & Formatting Rules:
- Picture: For the 'Picture' category, the text field should be omitted.
- Formula: Format its text as LaTeX.
- Table: Format its text as HTML.
- All Others (Text, Title, etc.): Format their text as Markdown.
4. Constraints:
- The output text must be the original text from the image, with no translation.
- All layout elements must be sorted according to human reading order.
5. Final Output: The entire output must be a single JSON object.""",
"πŸ” Solo rilevamento layout": """Please output only the layout detection information from the PDF image. For each detected element, output its bbox and category.
1. Bbox format: [x1, y1, x2, y2]
2. Layout Categories: ['Caption', 'Footnote', 'Formula', 'List-item', 'Page-footer', 'Page-header', 'Picture', 'Section-header', 'Table', 'Text', 'Title']
3. Output must be a single JSON array sorted by reading order.""",
"πŸ“ Solo testo (OCR)": """Please extract all text from the image. Output the text in reading order as plain text. Do not translate. Exclude page headers and footers.""",
"🌐 Web page parsing": """Please parse this web page screenshot. Extract all visible text content, UI elements, and their layout structure. Output as structured JSON.""",
"πŸ“ Scene text spotting": """Please detect and recognize all text visible in this scene/photo. For each text instance, output its bounding box and the recognized text. Output as JSON array.""",
}
def run_ocr(image: Image.Image, prompt_mode: str, max_tokens: int) -> tuple[str, str]:
if image is None:
return "⚠️ Carica un'immagine", ""
prompt = PROMPTS.get(prompt_mode, list(PROMPTS.values())[0])
tmp_path = "/tmp/dots_ocr_input.png"
image.save(tmp_path)
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": tmp_path},
{"type": "text", "text": prompt},
],
}
]
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
).to(DEVICE)
with torch.no_grad():
generated_ids = model.generate(**inputs, max_new_tokens=max_tokens)
generated_ids_trimmed = [
out_ids[len(in_ids):]
for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed,
skip_special_tokens=True,
clean_up_tokenization_spaces=False,
)[0]
# Formatta JSON
try:
parsed = json.loads(output_text)
json_output = json.dumps(parsed, indent=2, ensure_ascii=False)
except json.JSONDecodeError:
json_output = output_text
# Estrai markdown
markdown = ""
try:
parsed = json.loads(output_text) if isinstance(output_text, str) else output_text
if isinstance(parsed, list):
for item in parsed:
cat = item.get("category", "")
txt = item.get("text", "")
if cat == "Title":
markdown += f"# {txt}\n\n"
elif cat == "Section-header":
markdown += f"## {txt}\n\n"
elif cat == "Formula":
markdown += f"$$\n{txt}\n$$\n\n"
elif cat == "Table":
markdown += f"{txt}\n\n"
elif cat in ("Page-header", "Page-footer"):
continue
elif txt:
markdown += f"{txt}\n\n"
else:
markdown = output_text
except (json.JSONDecodeError, TypeError):
markdown = output_text
return json_output, markdown
def run_ocr_pdf(pdf_file, prompt_mode: str, max_tokens: int):
if pdf_file is None:
return "⚠️ Carica un PDF", ""
import fitz # PyMuPDF
doc = fitz.open(pdf_file.name)
all_json = []
all_md = []
for i, page in enumerate(doc):
pix = page.get_pixmap(dpi=200)
img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
j, m = run_ocr(img, prompt_mode, max_tokens)
all_json.append(f"// === Pagina {i + 1} ===\n{j}")
all_md.append(f"---\n### Pagina {i + 1}\n\n{m}")
doc.close()
return "\n\n".join(all_json), "\n".join(all_md)
# ── Interfaccia Gradio ──────────────────────────────────────────
with gr.Blocks(title="dots.ocr 1.5", theme=gr.themes.Soft()) as demo:
gr.Markdown("# πŸ” dots.ocr 1.5 β€” Document OCR")
gr.Markdown("Parsing multilingue di documenti con Vision-Language Model")
with gr.Row():
prompt_mode = gr.Dropdown(
choices=list(PROMPTS.keys()),
value="πŸ“„ Parsing completo (layout + testo)",
label="ModalitΓ ",
)
max_tokens = gr.Slider(
minimum=1000, maximum=24000, value=8000, step=1000,
label="Max tokens output",
)
with gr.Tabs():
with gr.TabItem("πŸ“· Immagine"):
with gr.Row():
img_input = gr.Image(type="pil", label="Carica immagine")
with gr.Column():
img_json = gr.Code(label="Output JSON", language="json")
img_md = gr.Markdown(label="Output Markdown")
img_btn = gr.Button("πŸš€ Analizza immagine", variant="primary")
img_btn.click(
fn=run_ocr,
inputs=[img_input, prompt_mode, max_tokens],
outputs=[img_json, img_md],
)
with gr.TabItem("πŸ“„ PDF"):
with gr.Row():
pdf_input = gr.File(label="Carica PDF", file_types=[".pdf"])
with gr.Column():
pdf_json = gr.Code(label="Output JSON", language="json")
pdf_md = gr.Markdown(label="Output Markdown")
pdf_btn = gr.Button("πŸš€ Analizza PDF", variant="primary")
pdf_btn.click(
fn=run_ocr_pdf,
inputs=[pdf_input, prompt_mode, max_tokens],
outputs=[pdf_json, pdf_md],
)
gr.Markdown("---\n*Powered by [dots.ocr 1.5](https://github.com/rednote-hilab/dots.ocr) su HF Spaces*")
if __name__ == "__main__":
demo.launch()