Preprocess the raw dataset in a form ReXPlug ingests. The raw dataset must be a JSON file with the same name as dataset_name and must be zipped in dataset_path. For example, AmazonDigitalMusic.json is zipped in the raw_datasets directory as AmazonDigitalMusic.zip. If you're using a non-Amazon dataset, please make sure that the JSON file has only four fields, namely, 'userId', 'itemId', 'review', 'rating', in that order. The truncate_after flag determines the number of interactions to be used for training the Discriminator.
python preprocess.py --dataset_name="AmazonDigitalMusic" --dataset_path="./data/raw_datasets/AmazonDigitalMusic.zip" --seed=1234 --truncate_after=100000
1. Training RRCA. For smaller datasets like Digital Music and Clothing, the training takes around 50-100 epochs. For larger datasets like BeerAdvocate and Yelp, training takes ~30 epochs. The code maintains a patience of 15 epochs.
python train_rrca.py --dataset_name="AmazonDigitalMusic" --batch_size_rrca=256 --learning_rate_rrca=0.002 --num_epochs_rrca=150
2. Training Discriminator. The following command trains the discriminator. 2-3 epochs usually suffice.
python train_discrim.py --dataset_name="AmazonDigitalMusic" --pretrained_model="gpt2-medium" --batch_size=64 --learning_rate=0.002 --epochs=3
The following command generates reviews from the trained ReXPlug.
python test.py --length=50 --num_iterations=1 --temperature=1 --sample --gamma=1 --gm_scale=0.875 --kl_scale=0.01 --num_reviews=5
@InProceedings{Hada21,
author = {Hada, Deepesh and M., Vijaikumar and Shevade, Shirish},
title = {ReXPlug: Explainable Recommendation Using Plug-and-Play Language Model},
booktitle = {Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval},
month = {July},
year = {2021}
}