Skip to content

Domain Generalization with MixStyle. ICLR'21.

Notifications You must be signed in to change notification settings

thucbx99/mixstyle-release

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

31 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MixStyle

This repo contains the code of our ICLR'21 paper, "Domain Generalization with MixStyle".

The OpenReview link is https://openreview.net/forum?id=6xHJ37MVxxp.

########## Updates ############

28-06-2021: A new implementation of MixStyle is out, which merges MixStyle2 to MixStyle and switches between random and cross-domain mixing using self.mix. The new features can be found here.

12-04-2021: A variable self._activated is added to MixStyle to better control the computational flow. To deactivate MixStyle without modifying the model code, one can do

def deactivate_mixstyle(m):
    if type(m) == MixStyle:
        m.set_activation_status(False)

model.apply(deactivate_mixstyle)

Similarly, to activate MixStyle, one can do

def activate_mixstyle(m):
    if type(m) == MixStyle:
        m.set_activation_status(True)

model.apply(activate_mixstyle)

Note that MixStyle has been included in Dassl.pytorch. See the code for details.

05-03-2021: You might also be interested in our recently released survey on domain generalization at https://arxiv.org/abs/2103.02503, which summarizes the ten-year development in domain generalization, with coverage on the history, datasets, related problems, methodologies, potential directions, and so on.

##############################

A brief introduction: The key idea of MixStyle is to probablistically mix instance-level feature statistics of training samples across source domains. MixStyle improves model robustness to domain shift by implicitly synthesizing new domains at the feature level for regularizing the training of convolutional neural networks. This idea is largely inspired by neural style transfer which has shown that feature statistics are closely related to image style and therefore arbitrary image style transfer can be achieved by switching the feature statistics between a content and a style image.

MixStyle is very easy to implement. Below we show a brief implementation of it in PyTorch. The full code can be found here.

import random
import torch
import torch.nn as nn


class MixStyle(nn.Module):
    """MixStyle.
    Reference:
      Zhou et al. Domain Generalization with MixStyle. ICLR 2021.
    """

    def __init__(self, p=0.5, alpha=0.1, eps=1e-6, mix='random'):
        """
        Args:
          p (float): probability of using MixStyle.
          alpha (float): parameter of the Beta distribution.
          eps (float): scaling parameter to avoid numerical issues.
          mix (str): how to mix.
        """
        super().__init__()
        self.p = p
        self.beta = torch.distributions.Beta(alpha, alpha)
        self.eps = eps
        self.alpha = alpha
        self.mix = mix
        self._activated = True

    def __repr__(self):
        return f'MixStyle(p={self.p}, alpha={self.alpha}, eps={self.eps}, mix={self.mix})'

    def set_activation_status(self, status=True):
        self._activated = status

    def update_mix_method(self, mix='random'):
        self.mix = mix

    def forward(self, x):
        if not self.training or not self._activated:
            return x

        if random.random() > self.p:
            return x

        B = x.size(0)

        mu = x.mean(dim=[2, 3], keepdim=True)
        var = x.var(dim=[2, 3], keepdim=True)
        sig = (var + self.eps).sqrt()
        mu, sig = mu.detach(), sig.detach()
        x_normed = (x-mu) / sig

        lmda = self.beta.sample((B, 1, 1, 1))
        lmda = lmda.to(x.device)

        if self.mix == 'random':
            # random shuffle
            perm = torch.randperm(B)

        elif self.mix == 'crossdomain':
            # split into two halves and swap the order
            perm = torch.arange(B - 1, -1, -1) # inverse index
            perm_b, perm_a = perm.chunk(2)
            perm_b = perm_b[torch.randperm(B // 2)]
            perm_a = perm_a[torch.randperm(B // 2)]
            perm = torch.cat([perm_b, perm_a], 0)

        else:
            raise NotImplementedError

        mu2, sig2 = mu[perm], sig[perm]
        mu_mix = mu*lmda + mu2 * (1-lmda)
        sig_mix = sig*lmda + sig2 * (1-lmda)

        return x_normed*sig_mix + mu_mix

How to apply MixStyle to your CNN models? Say you are using ResNet as the CNN architecture, and want to apply MixStyle after the 1st and 2nd residual blocks, you can first instantiate the MixStyle module using

self.mixstyle = MixStyle(p=0.5, alpha=0.1)

during network construction (in __init__()), and then apply MixStyle in the forward pass like

def forward(self, x):
    x = self.conv1(x) # 1st convolution layer
    x = self.res1(x) # 1st residual block
    x = self.mixstyle(x)
    x = self.res2(x) # 2nd residual block
    x = self.mixstyle(x)
    x = self.res3(x) # 3rd residual block
    x = self.res4(x) # 4th residual block
    ...

In our paper, we have demonstrated the effectiveness of MixStyle on three tasks: image classification, person re-identification, and reinforcement learning. The source code for reproducing all experiments can be found in mixstyle-release/imcls, mixstyle-release/reid, and mixstyle-release/rl, respectively.

Takeaways on applying MixStyle to your tasks:

  • Applying MixStyle to multiple lower layers is generally better
  • Do not apply MixStyle to the last layer that is the closest to the prediction layer
  • Different tasks might favor different combinations

For more analytical studies, please read our paper at https://openreview.net/forum?id=6xHJ37MVxxp.

To cite MixStyle in your publications, please use the following bibtex entry

@inproceedings{zhou2021mixstyle,
  title={Domain Generalization with MixStyle},
  author={Zhou, Kaiyang and Yang, Yongxin and Qiao, Yu and Xiang, Tao},
  booktitle={ICLR},
  year={2021}
}

About

Domain Generalization with MixStyle. ICLR'21.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 77.4%
  • C++ 20.9%
  • Other 1.7%