import gradio as gr import re from PIL import Image from io import BytesIO import torch from transformers import DonutProcessor, VisionEncoderDecoderModel def predict(inp): # 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") # 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 = "" 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 def upload_file(files): file_paths = [file.name for file in files] return file_paths gr.Interface(fn=predict, inputs=gr.Image(type="pil"), outputs="json").launch()