Torch-rnn

Not long ago I downloaded all the Messenger discussions I had with a friend. There were around 10,000 messages and I tough it would be fun to train an AI to learn these discussions, our way to write messages, and come up with discussions on its own.

To do that I use the Torch-rnn which is model of recurrent neural network designed to learn text. It’s base on Char-rnn

I developed a shell script to improve my workflow and make this wonderful tool more accessible.

See repo on github

Torch-rnn quickstart

built on the torch-rnn docker container.

Requirements

How to use

./torch-rnn-quickstart.sh <data.txt> [train, sample]
  • data.txt is the textyou want to feed the neural net for training

Training

./torch-rnn-quickstart.sh <data.txt> train

training parameters can be changes from the file training.conf generated in the corresponding folder (same path and name than the <data.txt>)

By default, the checkpoint with the most iterations will be used to continue training from.

Sampling

./torch-rnn-quickstart.sh <data.txt> sample

training parameters can be changes from the file sampling.conf generated in the corresponding folder (same path and name than the <data.txt>)

By default, the checkpoint with the most iterations will be used for sampling.