🚀 TrOCR for French
TrOCR has not yet released for French, so we trained a French model for PoC purpose. This model can be further enhanced by collecting more data.
🚀 Quick Start
TrOCR has not yet released for French, so we trained a French 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.
It's a special case of the English trOCR model introduced in the paper TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models by Li et al. and first released in this repository
This was possible thanks to daekun-ml and Niels Rogge than enabled us to publish this model with their tutorials and code.
✨ Features
Collecting data
Text data
We created training data of ~723k examples by taking random samples of the following datasets:
We collected parts of each of the datasets and then cut randomly the sentences to collect the final training set.
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 camembert-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 = 32
Results on dev set
For the dev set we got those results
- size of the test set: 72k examples
- CER: 0.13
- WER: 0.26
- Val Loss: 0.424
💻 Usage Examples
Basic Usage
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("agomberto/trocr-base-printed-fr")
tokenizer = AutoTokenizer.from_pretrained("agomberto/trocr-base-printed-fr")
url = "https://github.com/agombert/trocr-base-printed-fr/blob/main/sample_imgs/0.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=32)
generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(generated_text)
Additional Information
All the code required for data collection and model training has been published on the author's Github.
- https://github.com/agombert/trocr-base-printed-fr/
📄 License
This project is licensed under the MIT license.
%20--%3e%3cdefs%3e%3cstyle%3e%20.st0%20{%20fill:%20%23061b40;%20}%20.st1%20{%20fill:%20%23306af1;%20}%20.st2%20{%20fill:%20%235ce5cf;%20}%20%3c/style%3e%3c/defs%3e%3cg%3e%3cpath%20class='st0'%20d='M55,10.5h9v3h-9v9h12V7.5h-12v3ZM64,19.5h-6v-3h6v3Z'/%3e%3cpolygon%20class='st0'%20points='69%2016.5%2078%2016.5%2078%2019.5%2069%2019.5%2069%2022.5%2081%2022.5%2081%2013.5%2072%2013.5%2072%2010.5%2081%2010.5%2081%207.5%2069%207.5%2069%2016.5'/%3e%3cpolygon%20class='st0'%20points='95%2010.5%2095%207.5%2083%207.5%2083%2022.5%2095%2022.5%2095%2019.5%2086%2019.5%2086%2016.5%2095%2016.5%2095%2013.5%2086%2013.5%2086%2010.5%2095%2010.5'/%3e%3cpath%20class='st0'%20d='M40,1.5v21h11.6l1.4-1.4v-7.6h0c0,0-1.4-1.5-1.4-1.5l1.4-1.4V2.9l-1.4-1.4h-11.6ZM50,19.5h-7v-6h7v6ZM50,10.5h-7v-6h7v6Z'/%3e%3c/g%3e%3cpath%20class='st1'%20d='M23.1,24L14.7,4.8l-4.9,11.2h3.8l-1.8,4H3.3L12.1,0H2C.9,0,0,.9,0,2v20c0,1.1.9,2,2,2h21.1Z'/%3e%3cpath%20class='st2'%20d='M34,0h-16.8l10.6,24h6.2c1.1,0,2-.9,2-2V2C36,.9,35.1,0,34,0ZM32.5,20h-4V4h4v16Z'/%3e%3c/svg%3e)
