visirion07 / zeroshot Goto Github PK

0.) Data files are in ./data and baseline models are in ./models/baseline

1.) [Optional] Train the baseline classifier models for textual entailment: TE-embeddings and TE-lstm.

python train_baseline.py --data_path ./data --save_path ./models/baseline

You can also use the ones in ./models/baseline instead of training your own.

2.) Train the autoencoder (ARAE), generator and discriminator (GAN)

python train.py --data_path ./data --update_base --convolution_enc --classifier_path ./models

3.) Train the inverter

Once you have pretrained models for autoencoder, generator and discriminator, you can train the inverter as below

python train.py --data_path ./data --load_pretrained <pretrain_exp_ID> --classifier_path ./models

4.) By default, we use fast search instead of hybrid search (as described in the paper).

You can pass an argument --hybrid above to change that.

5.) TE-treeLSTM and machine translation results in the paper were done offline and are not included in the respository.

If you wish to get those results, you can contact [email protected] for further details.

Initial code is based on Zhao et al., 2017