|
import gradio as gr |
|
import re |
|
from PIL import Image |
|
from io import BytesIO |
|
import torch |
|
|
|
from transformers import DonutProcessor, VisionEncoderDecoderModel |
|
|
|
|
|
device = "cuda" if torch.cuda.is_available() else "cpu" |
|
|
|
|
|
processor = DonutProcessor.from_pretrained("jonathanjordan21/donut_fine_tuning_food_composition_id") |
|
|
|
|
|
model = VisionEncoderDecoderModel.from_pretrained("jonathanjordan21/donut_fine_tuning_food_composition_id") |
|
|
|
|
|
def predict(inp): |
|
|
|
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() |
|
|
|
return processor.token2json(sequence3) |
|
|
|
|
|
out = get_komposisi("", inp) |
|
return out |
|
|
|
|
|
|
|
gr.Interface(fn=predict, |
|
inputs=gr.Image(type="pil"), |
|
outputs="json").launch() |
|
|
