import gradio as gr
import re
from PIL import Image
from io import BytesIO
import torch

from transformers import DonutProcessor, VisionEncoderDecoderModel

# Check GPU
device = "cuda" if torch.cuda.is_available() else "cpu"
    
# Load processor
processor = DonutProcessor.from_pretrained("jonathanjordan21/donut_fine_tuning_food_composition_id")
    
# Load model
model = VisionEncoderDecoderModel.from_pretrained("jonathanjordan21/donut_fine_tuning_food_composition_id")


def predict(inp):
    # Define Json Parser
    def get_komposisi(image_path, image=None):
        image = Image.open(image_path).convert('RGB') if image== None else image.convert('RGB')
    
        task_prompt = "<s_kmpsi>"
        decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
    
        pixel_values = processor(image, return_tensors="pt").pixel_values
    
        outputs = model.generate(
            pixel_values.to(device),
            decoder_input_ids=decoder_input_ids.to(device),
            max_length=model.decoder.config.max_position_embeddings,
            early_stopping=True,
            pad_token_id=processor.tokenizer.pad_token_id,
            eos_token_id=processor.tokenizer.eos_token_id,
            use_cache=True,
            bad_words_ids=[[processor.tokenizer.unk_token_id]],
            return_dict_in_generate=True,
        )
    
        sequence1 = processor.batch_decode(outputs.sequences)[0]
        sequence2 = sequence1.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
        sequence3 = re.sub(r"<.*?>", "", sequence2, count=1).strip()  # remove first task start token
    
        return processor.token2json(sequence3)

    #Generate Output
    out = get_komposisi("", inp)
    return out



gr.Interface(fn=predict,
             inputs=gr.Image(type="pil"),
             outputs="json").launch()