simclr
is a reimplementation of "A Simple Framework for Contrastive Learning of Visual Representations" link
Example of minimum working command that launches a 2-gpu training with cifar10
python -m torch.distributed.launch --use_env --nproc_per_node=2 egg/zoo/simclr/train.py --batch_size=64 --dataset_name="cifar10" --dataset_dir="./cifar10" --image_size=32
In sweeps
there's a configuration that tries to reproduce the setup of the paper with a batch size of 2048 using 16 GPUs
In can be launched calling nest nest from the root directory of EGG with:
python egg/nest/nest.py --game egg.zoo.simclr.train --sweep egg/zoo/simclr/sweeps/simclr.json --checkpoint_dir="replicate_simclr" --nodes=2 --tasks=8