tf-char-rnn

This is a TensorFlow re-implementation of the char-rnn Lua code by Karpathy.

See Karpathy blog for description of what all this is about.

Running

You will need:

1. A text file with some data (you can use tinyshakespeare.txt from the data/ folder)
2. A machine with installed Python 3
3. (optional but highly recommended) a GPU on this

Configure and use virtual environment, install required packages:

virtualenv .venv -p python3
. .venv/bin/activate
pip install -r requirements.txt
pip install tensorflow  # or tensorflow-gpu, if you have GPU

Training

python train.py --input_data data/tinyshakespeare.txt --train_dir tinyshakespeare

Monitoring training and validation losses

Start TensorBoard server

python -m tensorflow.tensorboard.tensorboard --logdir `pwd`

Use web browser to connect to the TensorBoard server and see computational graph and plots of training and validation losses.

Sampling

python sample.py --load_model tinyshakespeare/lm_10000.model --prefix 'O thy!'