nabihach / IDA

Requirements for running the code:

1. NLTK
2. PyTorch0.4

Create an empty folder trained_models in the main directory. This is where trained models will be saved.

Download the MultiNLI (version 0.9) dataset, and put the jsonl files into the folder nli/data/raw/multinli/.

TRAINING ON SOURCE AND 1 TARGET

To train on a source domain (e.g. fiction), do python3 nli/nli_train.py -g fiction

The -g flag specifies which genre you want to train on. There are some other flags you can use, these can be found in nli/nli_train.py. The command above will run training and save the model in the folder trained_models/mnli_xxxxxxxx/, where xxxxxxx is a time-stamp.

For transfer learning on a different domain, e.g. government, run: python3 nli/nli_train.py -tl True -t mnli_xxxxxxx/_x.tar -tg government

where the -tl flag specifies that we wish to do transfer learning, -t specifies the model you want to load, and -tg is the target genre. The default number of slots is 500. To change this, you can change self.target_mem_slots in nli/encoder.py (line 29).

TESTING

To evaluate the model performance on the test set of a domain (e.g. government), follow the instructions given in comments of nli/encoder.py, make those changes, and then run: python3 nli/test_nli.py mnli_xxxxxxx/_x.tar government

TRAINING AND TEST ON MULTIPLE TARGET DOMAINS

Once you have trained on the source and ONE target domain, if you wish to train on a 2nd target domain, you need to change the following hard-coded settings in the code.

1. In nli/nli_train.py: Run model.encoder.add_target_pad_2() instead of model.encoder.add_target_pad() on line 95.
2. In nli/encoder.py: In __init__(), following the comments to make the necessary changes. Then follow the training and testing instructions given above.