Computational Intelligence Lab 2021

Project: Collaborative Filtering

Group name on Kaggle: Bayesians

Reproducing results

1. Install conda if it's not already installed

Download installer from here: https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

Give execution permission to the installer:

chmod +x Miniconda3-latest-Linux-x86_64.sh

Run the installer:

./Miniconda3-latest-Linux-x86_64.sh

Open a new terminal and verify that conda is installed.

2. Create the environment

conda env create -f environment.yml

3. Activate the environment

conda activate cil_project

4. Make the configurations in main.py

model should be "bayesian_gc_svdpp" or "gc_svdpp". Use "gc_svdpp" for Non-Bayesian SVD++ model which uses Graph Convolution Networks. Use "bayesian_gc_svdpp" for the Bayesian model which also calculates parameter and prediction uncertainties.

device should be "cuda" or "cpu". Use "cuda" if you will submit the job to a GPU.

data_path should be the path where the data is located.

save_dir should be the path where the predictions will be saved to.

5. Submit the training task to GPU with the following commands (indicated time is necessary to reproduce results)

cd src/

bsub -n 4 -W 24:00 -o logs -R "rusage[mem=4096, ngpus_excl_p=1]" -R "select[gpu_mtotal0>=10240]" python main.py

Predictions and uncertainty plots (if Bayesian mode is active) will be saved under the directory "save_dir".

Please refer to the README in src/baselines to reproduce the baseline results.

Some of this code was adapted from https://github.com/dmlc/dgl/tree/master/examples/pytorch/gcmc