Code heavily borrowed from InferSent repository.
- Dependencies of InferSent
- Comet.ML API
Use convert.py
to convert the data into InferSent repository format.
python train_nli.py --nlipath "/path/to/step_0/"
--outputdir "/path/to/outputs/"
--optimizer "adam,lr=0.001"
--encoder_type "InferSent"
--lrshrink 0.9
--enc_lstm_dim 300
--word_emb_type "normal"
--comet_apikey "<your api key>"
--comet_workspace "<your workspace>"
--comet_project "compositionality-nli"
--seed 20
score-perturbations.py
: Calculates statistics related to various probing perturbations on the output