README.md

Char RNN for creation of texts

Usage:

1. Clone this repository and make sure libraries from requirements.txt are installed
2. Prepare your text in .txt file
3. Open command line in repo
4. Command ./charrn.py train --filename /path/to/file.txt --verbose makes RNN to train (may take long cos it trains on cpu by default)
5. Command ./charnn.py generate --filename model.pt --start your_start_words_with_underscore --verbose makes the model to generate some text (200 symbols by default)

Train arguments:

--filename - path to text for training; required
--savepath - .pt filepath, where the model should be saved; default model.pt
--verbose - training with verbosity; default false
--model - lstm or gru options are available; default lstm
--batch_size - batch size for training; default 16
--seq_len - sequence len to train on; default 256
--hidden_size - hidden layers; default 128
--embedding_size - embedding size for symbols; default 128
--device - device to train on; cpu or cuda; default cpu
--layers - num of layers in RNN; default 2
--epochs - num of epochs; default 2000
--min_loss - min loss to train until; default 0.56 (to not to overfit)

Generate arguments:

--filename - /path/to/model.pt file
--model - lstm or gru; pass same as in training; default lstm
--verbose
--len - text len to generate; default 200
--start - start text to generate (with_underscores to parse)
--temp - temperature of text generation; the more, the more different from the original results should be observed; default 1.0 (no randomness)