A simple and complete implementation of Bootstrap your own latent: A new approach to self-supervised Learning in PyTorch + PyTorch Lightning.
- good performance (~67% linear eval accuracy on CIFAR100)
- minimal code, easy to use and extend
- multi-GPU / TPU and AMP support provided by PyTorch Lightning
- ImageNet support (needs testing)
- linear evaluation is performed during training without any additional forward pass
- logging with Wandb
Linear Evaluation Accuracy
Here is the accuracy after training for 1000 epochs:
|Dataset||[email protected]||[email protected]|
Training and Validation Curves
conda create --name essential-byol python=3.8 conda activate essential-byol conda install pytorch=1.7.1 torchvision=0.8.2 cudatoolkit=XX.X -c pytorch pip install pytorch-lightning==1.1.6 pytorch-lightning-bolts==0.3 wandb opencv-python
The code has been tested using these versions of the packages, but it will probably work with slightly different environments as well. When your run the code (see below for commands), PyTorch Lightning will probably throw a warning, advising you to install additional packages as
matplotlib. They are not needed for this implementation to work, but you can install them to get rid of the warnings.
Three datasets are supported:
For imagenet you need to pass the appropriate
--data_dir, while for CIFAR you can just pass
--download to download the dataset.
The repo comes with minimal model specific arguments, check
main.py for info. We also support all the arguments of the PyTorch Lightning trainer. Default parameters are optimized for CIFAR100 but can also be used for CIFAR10.
Sample commands for running CIFAR100 on a single GPU setup:
python main.py \ --gpus 1 \ --dataset CIFAR100 \ --batch_size 256 \ --max_epochs 1000 \ --arch resnet18 \ --precision 16 \ --comment wandb-comment
and multi-GPU setup:
python main.py \ --gpus 2 \ --distributed_backend ddp \ --sync_batchnorm \ --dataset CIFAR100 \ --batch_size 256 \ --max_epochs 1000 \ --arch resnet18 \ --precision 16 \ --comment wandb-comment
Logging is performed with Wandb, please create an account, and follow the configuration steps in the terminal. You can pass your username using
--entity. Training and validation stats are logged at every epoch. If you want to completely disable logging use
Help is appreciated. Stuff that needs work:
- test ImageNet performance
- exclude bias and bn from LARS adaptation (see comments in the code)