The codes for paper AutoTower: Efficient Neural Architecture Search for Two-Tower Recommendation

Taking ML-100k as an example.

• Download raw data of ml100k in https://grouplens.org/datasets/movielens/100k.
• Use codes/datasets/data/ml100k/process_ml100k.ipynb to preprocess the raw data.
• Use python codes/datasets/movielens100k.py to create dataset for further searching.

Make sure save locations of all datasets.

• Set the search space in codes/configs/config.py.
• Use sh scripts/train_supernet.sh to train the supernet, you need to set the data path in the scripts and some configurations in codes/train_supernet.py.

Use sh scripts/search_arch.sh to search the architectures, you need to set the data path in the scripts and some configurations in codes/train_search.py, especially the supernet tag.

The codes and models in this repo are released under the GNU GPLv3 license.