---
license: apache-2.0
language:
- ko
metrics:
- cer
- wer
pipeline_tag: image-to-text
---

# trOCR-youtube-kor-OCR

fine-tuned for VisionEncoderDecoderModel(encoder , decoder)
encoder = 'facebook/deit-base-distilled-patch16-384'
decoder = 'klue/roberta-base'

## How to Get Started with the Model

```python
from transformers import VisionEncoderDecoderModel,AutoTokenizer, TrOCRProcessor
import torch
from PIL import Image


device = torch.device('cuda') # change 'cuda' if you need.

image_path='(your image path)'
image = Image.open(image_path)
#model can be .jpg or .png
#hugging face download: https://huggingface.co/gg4ever/trOCR-final

processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
trocr_model = "gg4ever/trOCR-youtube-kor-OCR"
model = VisionEncoderDecoderModel.from_pretrained(trocr_model).to(device)
tokenizer = AutoTokenizer.from_pretrained(trocr_model)

pixel_values = (processor(image, return_tensors="pt").pixel_values).to(device)
generated_ids = model.generate(pixel_values)
generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(generated_text)

```

## Training Details
### Training Data

100k words generated by TextRecognitionDataGenerator(trdg) : https://github.com/Belval/TextRecognitionDataGenerator/blob/master/trdg/run.py
120k words from AI-hub OCR words dataset : https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&dataSetSn=81

### Training Hyperparameters

training_args = Seq2SeqTrainingArguments(
    predict_with_generate=True,
    evaluation_strategy="steps",
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    num_train_epochs=2,
    fp16=True,
    learning_rate=4e-5,
    output_dir="./models",
    save_steps=2000,
    eval_steps=1000,
    warmup_steps=2000,
    weight_decay=0.01
)