A PyTorch library for recognizing entailment between two sentences

Python 3.5+, PyTorch 0.3.0, CUDA 7.5+, tqdm

Default training uses the stanford SNLI dataset, you can download this dataset here. Put the snli_1.0 folder in the s2l repository.

You will also need to download the Stanford Glove word embeddings from here. Currently s2l only works with 300 dimensional word embeddings.

With defaults:

python main.py --cuda --expr_name=test_model --embed=/path/to/glove.6B.300d.txt

As of now, this package requires --cuda

--resume: path to the desired checkpoint file. Example -> --resume=exps/test_model/checkpoint.pth.tar

--h_size: size of LSTM hidden layer

--lstm_layers: number of layers in LSTM

--mlp_d: size of linear layer following LSTM

--vocab: default is vocabulary.txt, here you can create your own vocab

--labels: the entailment labels are saved in labels.txt, change for different labeling

--embed: path to glove.6B.300d.txt, as of now will only run with this glove embedding