Getting Started

This repository contains recognition modules for 8 different languages.

Language
Assamese
Bengali
Gujarati
Hindi
Marathi
Odia
Punjabi
Tamil

Versions

Version 1 at commit at 060efa6567a0ede9b0d08e14d922b488783548ac

Installation

Requires Python >= 3.9 and PyTorch >= 2.0. The default requirements files will install the latest versions of the dependencies (as of February 22, 2024).

Updating dependency version pins

conda create -n STocr python=3.9
pip install pip-tools
make clean-reqs reqs  # Regenerate all the requirements files

# Use specific platform build. Other PyTorch 2.0 options: cu118, cu121, rocm5.7
platform=cpu
# Generate requirements files for specified PyTorch platform
make torch-${platform}
# Install the project and core + train + test dependencies. Subsets: [train,test,bench,tune]
pip install -r requirements/core.${platform}.txt -e .[train,test]
pip install fire==0.6.0
pip install numpy==1.26.4

Inference

Following command is used to get inference on a set of images from desired model options available in assets.

python your_script.py \
--checkpoint /path/to/checkpoint.ckpt \
--language hindi \
--image_dir /path/to/images \
--save_dir /path/to/save

For english you can directly run python infer.py --checkpoint --image_dir images --language english --save_dir output To check the argument usage python infer.py --help.

Acknowledgments

Text Recognition - PARseq