ABOUT = """
# TB-OCR Preview 0.1 Unofficial Demo

This is an unofficial demo of [yifeihu/TB-OCR-preview-0.1](https://huggingface.co/yifeihu/TB-OCR-preview-0.1).

Overview of TB-OCR:

> TB-OCR-preview (Text Block OCR), created by [Yifei Hu](https://x.com/hu_yifei), is an end-to-end OCR model handling text, math latex, and markdown formats all at once. The model takes a block of text as the input and returns clean markdown output. Headers are marked with `##`. Math expressions are guaranteed to be wrapped in brackets `\( inline math \) \[ display math \]` for easier parsing. This model does not require line-detection or math formula detection. 

(From the [model card](https://huggingface.co/yifeihu/TB-OCR-preview-0.1))
"""
# check out https://huggingface.co/microsoft/Phi-3.5-vision-instruct for more details

import torch, spaces
from transformers import AutoModelForCausalLM, AutoProcessor
from PIL import Image
import requests
import os
# os.system('pip install -U flash-attn')
model_id = "yifeihu/TB-OCR-preview-0.1"

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

if not torch.cuda.is_available():
    ABOUT += "\n\n### ⚠️ This demo is running on CPU ⚠️\n\nThis demo is running on CPU, it will be very slow. Consider duplicating it or running it locally to skip the queue and for faster response times."

model = AutoModelForCausalLM.from_pretrained(
  model_id, 
  device_map=DEVICE, 
  trust_remote_code=True, 
  torch_dtype="auto",
  # _attn_implementation='flash_attention_2',
  #load_in_4bit=True # Optional: Load model in 4-bit mode to save memory
)

processor = AutoProcessor.from_pretrained(model_id, 
  trust_remote_code=True, 
  num_crops=16
)
@spaces.GPU
def phi_ocr(image_url):
    question = "Convert the text to markdown format."
    image = Image.open(image_url)
    prompt_message = [{
        'role': 'user',
        'content': f'<|image_1|>\n{question}',
    }]

    prompt = processor.tokenizer.apply_chat_template(prompt_message, tokenize=False, add_generation_prompt=True)
    inputs = processor(prompt, [image], return_tensors="pt").to(DEVICE) 

    generation_args = { 
        "max_new_tokens": 1024, 
        "temperature": 0.1, 
        "do_sample": False
    }

    generate_ids = model.generate(**inputs, eos_token_id=processor.tokenizer.eos_token_id, **generation_args
    )

    generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
    response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] 

    response = response.split("<image_end>")[0] # remove the image_end token 

    return response

import gradio as gr

with gr.Blocks() as demo:
    gr.Markdown(ABOUT)
    with gr.Row():
        with gr.Column():
            img = gr.Image(label="Input image", type="filepath")
            btn = gr.Button("OCR")
        with gr.Column():
            out = gr.Markdown()
    btn.click(phi_ocr, inputs=img, outputs=out)
demo.queue().launch()