BYOL-PyTorch

PyTorch implementation of Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning with DDP (DistributedDataParallel) and Apex Amp (Automatic Mixed Precision).

Requirements

python>=3.6.9
pytorch>=1.4.0
tensorboardx
opencv-python==4.2.0.34
pyyaml==5.3.1
apex

This repo supposes using torch.distributed.launch to start training, for example:

python -m torch.distributed.launch --nproc_per_node=2 --nnodes=2 --node_rank=0 --master_addr="" --master_port=12345 byol_main.py

Implementation Details

1. Use apex or pytorch>=1.4.0 for SyncBatchNorm
2. Pay attention to the data augmentations, which are slightly different from those in SimCLR, especially the probability of applying GaussianBlur and Solarization for different views (see Table 6 of the paper)
3. In both training and evaluation, they normalize color channels by subtracting the average color and dividing by the standard deviation, computed on ImageNet, after applying the augmentations (even with the specially designed augmentations)
4. Increase target model momentum factor with a cosine rule
5. Exclude biases and batch normalization parameters from both LARS adaptation and weight decay
6. The correct order for model wrapping: convert_syncbn -> cuda -> amp.initialize -> DDP

Reproduced Results

Linear Evaluation Protocol

Here we post our reproduced results with hyper parameters in train_config.yaml using 32x Nvidia V100 (32GB) GPU cards, indicating a global batch size of 4096.

Under this setup, reference accuracies for 300 epochs are 72.5% (top-1) and 90.8% (top-5), as reported in Section F of the paper.

Train Epoch	Classifier Train Epoch	Classifier LR	Top-1 ACC	Top-5 ACC
100	100	[1., 0.]/Cosine	64.34%	85.95%
200	100	[1., 0.]/Cosine	69.48%	89.07%
300	100	[1., 0.]/Cosine	72.29%	90.90%

Model Checkpoints

We release our BYOL checkpoint of ResNet-50 model trained on ImageNet (ILSVRC 2012) for 300 epochs on Google Drive.

Note: This checkpoint file is created using legacy codebase, which saves the online_backbone, online_projection, predictor and target network modules separately. And the parameter names are slightly different from ResNet50 in torchvision.models. Concretely, parameters in checkpoint['online_backbone'] is probably what you want, from conv1 to avgpool, yielding 318 parameters.

You can refer to utils/load_and_convert.py for the parameter name agnostic loading and converting (to torchscript).

Remarks

I have modified this repo (remove some code) to meet the compliance issues from my former employer. If you have forked or saved the legacy versions, please delete them.