import torch
import re
from PIL import Image
from transformers import DonutProcessor, VisionEncoderDecoderModel
import base64
import io
import json

def model_fn(model_dir):
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    processor = DonutProcessor.from_pretrained(model_dir)
    model = VisionEncoderDecoderModel.from_pretrained(model_dir)
    model.to(device)
    return model, processor, device

def transform_fn(model, request_body, input_content_type, output_content_type):
    model, processor, device = model
    if input_content_type == 'application/json':
        data = json.loads(request_body)
        image_data = data['inputs']
        image = Image.open(io.BytesIO(base64.b64decode(image_data))).convert("RGB")
        
        # Preprocess the image
        pixel_values = processor(image, return_tensors="pt").pixel_values.to(device)
        
        # Run inference
        model.eval()
        with torch.no_grad():
            task_prompt = "<s_receipt>"
            decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids.to(device)
            generated_outputs = model.generate(
                pixel_values,
                decoder_input_ids=decoder_input_ids,
                max_length=model.config.decoder.max_position_embeddings, 
                pad_token_id=processor.tokenizer.pad_token_id,
                eos_token_id=processor.tokenizer.eos_token_id,
                early_stopping=True,
                bad_words_ids=[[processor.tokenizer.unk_token_id]],
                return_dict_in_generate=True
            )
        
        # Decode the output
        decoded_text = processor.batch_decode(generated_outputs.sequences)[0]
        decoded_text = decoded_text.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
        decoded_text = re.sub(r"<.*?>", "", decoded_text, count=1).strip()
        
        # Prepare the response
        prediction = {'result': decoded_text}
        return json.dumps(prediction), output_content_type
    else:
        raise ValueError(f"Unsupported content type: {input_content_type}")