Shuffle BatchNorm
Note: please refer to github.com/facebookresearch/moco for official implementation. This implementation is not optimal.
An implementation of Shuffle BatchNorm technique mentioned in He et al., Momentum Contrast for Unsupervised Visual Representation Learning, 2019, in Section 3.3 "Shuffling BN".
Implemented with torch 1.3.1. It works with pytorch DistrbutedDataParallel with 1 process per GPU. So in order to use this ShuffleBatchNorm layer you need at least 2 GPUs.
What's this?
The formula above is the BatchNorm algorithm. The ShuffleBatchNorm layer shuffles the mini-batch statistics (mean and variance) across multiple GPUs to avoid information leak. This operation eliminates model "cheating" when training contrastive loss and the contrast is obtained within the mini batch.
How to use?
The implementation mimics the design of SyncBatchNorm. To use ShuffleBatchNorm, just create your model first and then convert all torch.nn.BatchNormND layers into ShuffleBatchNorm by the function:
from shuffle_batchnorm import ShuffleBatchNorm
# ...
model = Model() # with BN layers
model = ShuffleBatchNorm.convert_shuffle_batchnorm(model)See main.py for a completed example.
Check result
run command:
$ python main.py --gpu 0,1 --shuffle --epochs 10
=> Spawning 2 distributed workers
...
[0]mean before shuffle: tensor([-0.2478, 0.1704, 0.0640, -0.2732], device='cuda:0')
[1]mean before shuffle: tensor([-0.4012, -0.1913, -0.0553, -0.1917], device='cuda:1')
[0]mean after shuffle: tensor([-0.4012, -0.1913, -0.0553, -0.1917], device='cuda:0')
[1]mean after shuffle: tensor([-0.2478, 0.1704, 0.0640, -0.2732], device='cuda:1')
[9/10] Loss 0.6868
================================================
[9/10] Loss 0.7908
================================================Notes
If you find bugs, please create an issue. Very welcome!
Update:
- Doesn't work when training with multiple nodes, will fix soon.