Skip to content

Standardizing weights to accelerate micro-batch training

Notifications You must be signed in to change notification settings

LoserSun/WeightStandardization

 
 

Repository files navigation

Weight Standardization

Weight Standardization (WS) is a normalization method to accelerate micro-batch training. Micro-batch training is hard because small batch sizes are not enough for training networks with Batch Normalization (BN), while other normalization methods that do not rely on batch knowledge still have difficulty matching the performances of BN in large-batch training.

Our WS ends this problem because when used with Group Normalization and trained with 1 image/GPU, WS is able to match or outperform the performances of BN trained with large batch sizes with only 2 more lines of code. So if you are facing any micro-batch training problem, please do yourself a favor and try Weight Standardization. You will be surprised by how well it performs.

WS achieves these superior results by standardizing the weights in the convolutional layers, which we show is able to smooth the loss landscape by reducing the Lipschitz constants of the loss and the gradients. Please see our arXiv report for the details. If you find this project helpful, please consider citing our paper.

@article{weightstandardization,
  author    = {Siyuan Qiao and Huiyu Wang and Chenxi Liu and Wei Shen and Alan Yuille},
  title     = {Weight Standardization},
  journal   = {arXiv preprint arXiv:1903.10520},
  year      = {2019},
}

Weight Standardization on computer vision tasks

This repo holds our implementations of Weight Standardization for the following tasks.

Task Folder
ImageNet classification pytorch-classification
Object detection and instance segmentation on COCO maskrcnn-benchmark
Semantic segmentation on PASCAL VOC DeepLabv3.pytorch
Point cloud classification on ModelNet40 dgcnn

Implementing WS as a layer

We provide layer implementation of WS in TensorFlow and PyTorch. Replacing the convolutional layers with the following ones will give you performance boosts for free. NOTE: only replace convolutional layers that are followed by normalization layers such as BN, GN, etc.

PyTorch

class Conv2d(nn.Conv2d):

    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
                 padding=0, dilation=1, groups=1, bias=True):
        super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride,
                 padding, dilation, groups, bias)

    def forward(self, x):
        weight = self.weight
        weight_mean = weight.mean(dim=1, keepdim=True).mean(dim=2,
                                  keepdim=True).mean(dim=3, keepdim=True)
        weight = weight - weight_mean
        std = weight.view(weight.size(0), -1).std(dim=1).view(-1, 1, 1, 1) + 1e-5
        weight = weight / std.expand_as(weight)
        return F.conv2d(x, weight, self.bias, self.stride,
                        self.padding, self.dilation, self.groups)

TensorFlow

kernel_mean = tf.math.reduce_mean(kernel, axis=[0, 1, 2], keepdims=True, name='kernel_mean')
kernel = kernel - kernel_mean
#kernel_std = tf.math.reduce_std(kernel, axis=[0, 1, 2], keepdims=True, name='kernel_std')
kernel_std = tf.keras.backend.std(kernel, axis=[0, 1, 2], keepdims=True)
kernel = kernel / (kernel_std + 1e-5)

About

Standardizing weights to accelerate micro-batch training

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published