README.md

metadata

widget:
  - text: >-
      def add ( severity , progname , & block ) return true if io . nil? ||
      severity < level message = format_message ( severity , progname , yield )
      MUTEX . synchronize { io . write ( message ) } true end
license: mit
language:
  - id
  - en
pipeline_tag: document-question-answering

Model description

This model is based on the naver-clova-ix/donut-base model. The training dataset is created by manually scrapping images across the internet

Usage & limitations

The model could be used to detect the nutritional facts or compositions from images of food or drug packages. It is capable to create a json format of the components described in the image. However, due to lack of data, the texts in the image must be concisely upright.

Output Example

Model Output :

'<s_kmpsi><s_komposisi><s_obat>Vitamin E</s_obat><s_takaran>30 I.U.</s_takaran><sep/><s_obat>Tiamin HCl (B1)</s_obat><s_takaran>100 mg</s_takaran><sep/><s_obat>Piridoksin HCl (B6)</s_obat><s_takaran>50 mg</s_takaran><sep/><s_obat>Sianokobalamin (B12)</s_obat><s_takaran>100 mcg</s_takaran><sep/><s_obat>K-l-aspartat</s_obat><s_takaran>100 mg</s_takaran><sep/><s_obat>Mg-l-aspartat</s_obat><s_takaran>100 mg</s_takaran></s_komposisi><s_desc></s_desc></s_kmpsi>'

Json Parsed Output :

{'komposisi': [{'obat': 'Vitamin E', 'takaran': '30 I.U.'}, {'obat': 'Tiamin HCl (B1)', 'takaran': '100 mg'}, {'obat': 'Piridoksin HCl (B6)', 'takaran': '50 mg'}, {'obat': 'Sianokobalamin (B12)', 'takaran': '100 mcg'}, {'obat': 'K-l-aspartat', 'takaran': '100 mg'}, {'obat': 'Mg-l-aspartat', 'takaran': '100 mg'}], 'desc': ''}

How to use

Load Donut Processor and Model

import re
from transformers import DonutProcessor, VisionEncoderDecoderModel

# 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")

Create JSON parser

from PIL import Image
from io import BytesIO
import torch

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)

Get json output from an image

import requests
import io

image = requests.get('https://pintarjualan.id/wp-content/uploads/sites/2/2022/04/label-nustrisi-fact-1.png').content
print(get_komposisi("", Image.open(io.BytesIO(image))))