vladbataev / handwriting_synthesis

Handwriting synthesis with deep sequence models. Reimplementation of the paper https://arxiv.org/abs/1308.0850

Handwriting synthesis

Generation of handwriting text with deep neural networks. Reimplementation of the paper https://arxiv.org/abs/1308.0850 using pytorch.

Installation

Clone the repository and run pip3 install .

Model training

You can train your own models running scripts bin/train_prediction.sh or bin/train_synthesis.sh. Pretrained models can be found in pretrained folder.

Demonstration notebook

The generation and prediction results can be found in notebooks/results.ipynb

Some training tips

To get attention diagonal much faster it's useful to calculate average fraction of out and in sequence lengths and introduce new multiplication coefficient $\alpha$ for the update of attention values (p.26 formula(51) original paper):