To train:

> python char-rnn.pytorch/train.py --train_set data/train.txt --valid_set data/valid.txt --model lstm --cuda

To generate text (use —-cpu to use the GPU-trained model on a CPU):

> cd char-rnn.pytorch
> python generate.py all_tex_files.pt  -t 0.9 -l 500 --prime_str “The model " --cpu

The weights are saved as gru_epoch1_nlayers3_input100_output100_hs200_trainL20.3_valL30.2.pt, where

• gru is the model type (can also be lstm)
• epoch is the final epoch (i.e. max_epoch or when validation bpc started to augment)
• nlayers is the number of layers
• input is the input size
• output is the output size
• hs is the size of the hidden layers
• trainL is the loss on the train set
• valL is the loss on the validation set

It's an ugly but descriptive name.