alexholdenmiller / bagel

llvm project

basics

# for config and plotting
pip installd -r requirements.txt
python main.py

# for logging
wandb login
python main.py wandb_log=True

For set up on greene check out this: https://github.com/alexholdenmiller/nyu_cluster/edit/main/README.md

# to run a single job on greene
# srun...
python main.py

# to run a sweep (don't srun first)
python main.py -m lr=0.1,0.01 batch_size=128,32

Reproducing patch embedding, convolutions, and vector quantization

python main.py model=vit
python main.py model=convt
python main.py model=vqt

For reproducing the vq_vae encoding + vq + vit, patch vit and conv + vq + vit experiments

python vqvae.py # to train the vqvae model

python vit_vqvae.py # to train vit with vqvae encoding (specify modes: conv, patch, vqvae for comparisons)

We provide 2 notebook demos: vit-cifar-10.ipynb and vq-vae.ipynb to deonstrate base vit and vqvae runs.