A clear and maintainable implementation of Attention OCR in Tensorflow 2.0.

This sequence to sequence OCR model aims to provide a clear and maintainable implementation of attention based OCR.

Please note that this is currently a work in progress. Documentation is missing, but will be added when the code is stable.

This repository depends upon the following:

• Tensorflow 2.0
• Python 3.6+

To train a model, first download the sources for generating synthetic data:

cd synthetic
./download_data_sources.sh

Next, in this project's root folder, run the training script:

python3 run.py

This will run a test training run. If everything went well, you'll find a file named "trained.h5" in your directory. To train a real model you should change the training parameters. See run.py its arguments to find out what is configurable.

python3 run.py --help

This work is based on the following work:

• Attention Is All You Need

• Make image height variable
• Name all input and output tensors
• Write unit tests with full coverage
• Show a test case on google colab
• Perform a grid search on best parameters for a toy dataset
• Document the whole API