DFA extractor

Extracting DFAs from RNNs using learning and state merging techniques.

Git LFS

This project uses Git LFS to track .th files for models. See here for more info.

Command Reference

To train an RNN language model:

python train_rnn.py

Tomita 6/7 should both reach 100% dev accuracy within 2 epochs on the default seed:

python train_rnn.py --lang=Tom6 --n_train=100000

To extract a DFA from an RNN, and create a plot of accuracy vs number of data in {1, ..., 20} (doesn't require openfst):

python extract_dfa.py --lang=Tom6 --n_train_low=2 --n_train_high=20

To train an RNN for 100 epochs and save all checkpoints:

python train_rnn.py --save_name=Tom7-100 --lang=Tom7 --save_all --n_epochs=100 --stop_threshold=1000