CARD: Non-Uniform Quantization of Visual Semantic Unit for Generative Recommendation

Framework

🛠️ Environment

python==3.8.13
torch==2.0.0
torchvision==0.15.1
torchaudio==2.0.1
transformers==4.46.2
accelerate==0.8.0

📦 Dataset

We use the Amazon Reviews 2014 dataset. You may download it.

cd process_data

📁 Processes the raw dataset.

python process_data.py

🔧 Constructs interaction sequences for SASRec.

python process_collaborative_data.py

Then the SASRec model is employed to capture sequential user behaviors and derive product embeddings.

▶️ Construct visual semantic units.

python compose_card.py

⚙️ Encodes visual semantic units.

python encoder_card.py

🚀 Training Pipeline

train the NU-RQ-VAE

cd nu-rq-vae
python main.py
python generate_code.py

train the model

cd model
python main.py

Acknowledgement

We thank TIGER for providing the useful source code.