PyTorch implementation of char-rnn (character-level language model)
Download this Shakespeare dataset (from the original char-rnn) as
shakespeare.txt. Or bring your own dataset — it should be a plain text file (preferably ASCII).
train.pywith the dataset filename to train and save the network:
> python train.py shakespeare.txt
Training for 2000 epochs... (... 10 minutes later ...) Saved as shakespeare.pt
After training the model will be saved as
Usage: train.py [filename] [options]
Options: --model Whether to use LSTM or GRU units gru --n_epochs Number of epochs to train 2000 --print_every Log learning rate at this interval 100 --hidden_size Hidden size of GRU 50 --n_layers Number of GRU layers 2 --learning_rate Learning rate 0.01 --chunk_len Length of training chunks 200 --batch_size Number of examples per batch 100 --cuda Use CUDA
generate.pywith the saved model from training, and a "priming string" to start the text with.
> python generate.py shakespeare.pt --prime_str "Where"
Where, you, and if to our with his drid's Weasteria nobrand this by then.
AUTENES: It his zersit at he
Usage: generate.py [filename] [options]
Options: -p, --prime_str String to prime generation with -l, --predict_len Length of prediction -t, --temperature Temperature (higher is more chaotic) --cuda Use CUDA