Reproducibility Guide
Overview
This part of repo contains the implementation and experiments. This guide will help you reproduce the results using Docker or manual installation.
Docker Setup (Recommended)
1. Build Docker Image
docker build -t yambda-image .
2. Run Container with GPU Support
docker run --gpus all \
--runtime=nvidia \
-it \
-v </absolute/path/to/local/data>:/yambda/data \
yambda-image
Data Organization
Create following structure in mounted data directory:
data/
βββ flat/
β βββ 50m/
β βββ likes.parquet
β βββ listens.parquet
β βββ ...
βββ sequential/
βββ 50m/
βββ likes.parquet
βββ listens.parquet
βββ ...
Note: Sequential data is only needed for sasrec. You can build it from flat using scripts/transform2sequential.py or download
Running Experiments
General Usage
# For example random_rec
cd models/random_rec/
# Show help for main script
python main.py --help
# Basic execution
python main.py
Specific Methods
BPR/ALS
cd models/bpr_als
python main.py --model bpr
python main.py --model als
SASRec
cd models/sasrec
# Training
python train.py --exp_name exp1
# Evaluation
python eval.py --exp_name exp1
Manual Installation (Not Recommedned)
1. Install Core Dependencies
pip install torch torchvision torchaudio
2. Install Implicit (CUDA 11.8 required)
Implicit works only with cuda<12. See reasons here
CUDACXX=/usr/local/cuda-11.8/bin/nvcc \
pip install implicit
3. Install SANSA
sudo apt-get install libsuitesparse-dev
git clone https://github.com/glami/sansa.git
cd sansa && \
SUITESPARSE_INCLUDE_DIR=/usr/include/suitesparse \
SUITESPARSE_LIBRARY_DIR=/usr/lib \
pip install .
4. Install Project Package
pip install .