daekeun-ml's picture
Update README.md
b8e0ab4
metadata
language:
  - ko
tags:
  - trocr
  - image-to-text
license: mit
metrics:
  - wer
  - cer
widget:
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/random_2.jpg
    example_title: 랜덤 문장 1
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/random_6.jpg
    example_title: 랜덤 문장 2
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/chatbot_3.jpg
    example_title: 챗봇 1
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/chatbot_5.jpg
    example_title: 챗봇 2
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/news_1.jpg
    example_title: 뉴스 1
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/news_3.jpg
    example_title: 뉴스 2
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/nsmc_1.jpg
    example_title: 영화 리뷰 1
  - src: >-
      https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/nsmc_2.jpg
    example_title: 영화 리뷰 2

TrOCR for Korean Language (PoC)

Overview

TrOCR has not yet released a multilingual model including Korean, so we trained a Korean model for PoC purpose. Based on this model, it is recommended to collect more data to additionally train the 1st stage or perform fine-tuning as the 2nd stage.

Collecting data

Text data

We created training data by processing three types of datasets.

For efficient data collection, each sentence was separated by a sentence separator library (Kiwi Python wrapper; https://github.com/bab2min/kiwipiepy), and as a result, 637,401 samples were collected.

Image Data

Image data was generated with TextRecognitionDataGenerator (https://github.com/Belval/TextRecognitionDataGenerator) introduced in the TrOCR paper. Below is a code snippet for generating images.

python3 ./trdg/run.py -i ocr_dataset_poc.txt -w 5 -t {num_cores} -f 64 -l ko -c {num_samples} -na 2 --output_dir {dataset_dir}

Training

Base model

The encoder model used facebook/deit-base-distilled-patch16-384 and the decoder model used klue/roberta-base. It is easier than training by starting weights from microsoft/trocr-base-stage1.

Parameters

We used heuristic parameters without separate hyperparameter tuning.

  • learning_rate = 4e-5
  • epochs = 25
  • fp16 = True
  • max_length = 64

Usage

inference.py

from transformers import TrOCRProcessor, VisionEncoderDecoderModel, AutoTokenizer
import requests 
from io import BytesIO
from PIL import Image

processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten") 
model = VisionEncoderDecoderModel.from_pretrained("daekeun-ml/ko-trocr-base-nsmc-news-chatbot")
tokenizer = AutoTokenizer.from_pretrained("daekeun-ml/ko-trocr-base-nsmc-news-chatbot")

url = "https://raw.githubusercontent.com/aws-samples/sm-kornlp/main/trocr/sample_imgs/news_1.jpg"
response = requests.get(url)
img = Image.open(BytesIO(response.content))

pixel_values = processor(img, return_tensors="pt").pixel_values 
generated_ids = model.generate(pixel_values, max_length=64)
generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] 
print(generated_text)

All the code required for data collection and model training has been published on the author's Github.