benchmarks/README.md

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/
├── flat/
│   └── 50m/
│       ├── likes.parquet
│       ├── listens.parquet
│       └── ...
└── sequential/
    └── 50m/
        ├── likes.parquet
        ├── listens.parquet
        └── ...

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 .