This repository contains my attempt at implementing Alex Graves wonderful paper on sequence generation in Pytorch.
Most of the data pre-processing and sampling code is from hardmaru's implementation. Since these parts are independent of the framework, can be used directly.
Tried to keep the model and loss function implementation simple to follow. Hope it helps anybody looking for a simple implementation.
conda install -c omnia svgwrite
sudo apt-get install libmagickwand-dev
pip install Wand
pip install tensorboardX
pytorch >= 0.3
Before running train.py and sample.py you need to follow the instruction and download the necessary files.
You can find a model trained with the default parameters for 30 epochs in the save directory. Below is the training curve for the trained model.
Some generated samples from the trained network:
Well it does look like handwriting!