chainer-char-rnn

karpathy's char-rnn implementation by Chainer

Requirement

• Chainer

$ pip install chainer
$ pip install chainer-cuda-deps

Train

Start training the model using train.py, for example

$ python train.py

The --data_dir flag specifies the dataset to use. By default it is set to data/tinyshakespeare which consists of a subset of works of Shakespeare.

Your own data: If you'd like to use your own data create a single file input.txt and place it into a folder in data/. For example, data/some_folder/input.txt.

Sampling

Given a checkpoint file (such as those written to cv) we can generate new text. For example:

$ python sample.py \
--vocabulary data/tinyshakespeare/vocab.bin \
--model cv/some_checkpoint.chainermodel \
--primetext some_text --gpu -1

References

• Original implementation: https://github.com/karpathy/char-rnn
• Blog post: http://karpathy.github.io/2015/05/21/rnn-effectiveness/