Symmetry Structured Convolutional Neural Networks

The code is tested on a single NVIDIA® QUADRO® P5000 with PyTorch 1.1.0 and Python 3.6.9.

Requirements

pip install -r requirements.txt

Implementation

SCosRec convolutional kernels use the half Normal Glorot initialization.

Training SCosRec models from scratch

To train SCosRec model on ml1m dataset with learning rate decay (best results):

python train.py --dataset=ml1m --lr_decay True --model_type=scnn

Note: You should be able to obtain MAPs of ~0.1970 with the above settings.

To train SCosRec model on gowalla dataset with learning rate decay (best results):

python train.py --dataset=gowalla --d=100 --fc_dim=50 --l2=1e-6 --lr_decay True --model_type=scnn

Note: You should be able to obtain MAPs of ~0.1006 with the above settings.

Acknowledgments

This project is built on top of CosRec.