MABN的学习笔记--Towards Stabilizing Batch Statistics in Backward Propagation of Batch Normalization

Towards Stabilizing Batch Statistics in Backward Propagation of Batch Normalization

MABN是一种新颖的归一化方法，性能优于SynCBN、BRN、BN等方法。作者团队旷视科技&复旦，该方法一小时前开源！
点击获取论文https://openreview.net/forum?id=SkgGjRVKDS&noteId=BJeCWt3KiH，github链接https://github.com/megvii-model/MABN。

这里只简单介绍MABN和它的使用。

论文摘要

BN是深度学习领域中使用最广泛的技术之一。但是由于批次大小不足，其性能可能会严重下降。这种弱点限制了BN在许多计算机视觉任务上的使用，由于内存消耗的限制，批处理的大小通常很小。因此，已经提出了许多改进的归一化技术，它们要么不能完全恢复BN的性能，要么必须在推理过程中引入额外的非线性运算并增加巨大的消耗。
本文中，我们揭示了BN向后传播涉及两个额外的批处理统计信息。与梯度相关的额外批处理统计信息会严重影响深度神经网络的训练，根据分析，我们提出了一种新颖的归一化方法，称之为移动平均批量归一化（MABN）。在小批量情况下，MABN可以完全恢复vanilla BN的性能，而无需在推理过程中引入任何其它非线性操作。我们通过理论分析和实验证明了MABN的好处。通过实验，我们证明了MABN在包括Imagenet和COCO在内的多种计算机视觉任务的有效性。

实验结果

由上图可以看出，在Batch Size很小的情况下，MABN也得到了很好的效果。

MABN在batch_size很小等于2的时候，对于Imagenet数据集，相比BN和BRN，得到的Top 1 Accuracy与使用BN且batch_size=32时相当。

对于COCO数据集，MABN仅在batch_size=2就可获得了其它归一化方法需要将batch_size设置为16才能获得的精度，进一步说明了MABN在小批量数据下的有效性和优良性。

MABN的简单使用

以resnet50为例，将MABN替代resnet50网络构建用到的nn.BatchNorm2d。
将MABN.py文件包括在工作目录，
重点注意from networks.MABN import MABN2d as BatchNorm2d
resnet.py中的nn.BatchNorm2d用BatchNorm2d代替，改写为：

import torch
import torch.nn as nn
import torch.nn.functional as F

from networks.MABN import MABN2d as BatchNorm2d


class Conv2d(nn.Module):
    """Conv2d layer with Weight Centralization
    """
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1,
                 padding=0, dilation=1, groups=1, bias=False):
        super(Conv2d, self).__init__()
        self.in_planes = in_planes
        self.out_planes = out_planes
        self.stride = stride
        self.padding = padding
        self.dilation = dilation
        self.groups = groups
        self.weight = nn.Parameter(torch.randn(out_planes, in_planes//groups, kernel_size, kernel_size))
        if bias:
            self.bias = nn.Parameter(torch.randn(out_planes))
        else:
            self.register_parameter('bias', None)

    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
        return F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups)


def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)


def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)


class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = BatchNorm2d(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = conv1x1(inplanes, planes)
        self.bn1 = BatchNorm2d(planes)
        self.conv2 = conv3x3(planes, planes, stride)
        self.bn2 = BatchNorm2d(planes)
        self.conv3 = conv1x1(planes, planes * self.expansion)
        self.bn3 = BatchNorm2d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out


class ResNet(nn.Module):

    def __init__(self, block, layers, num_classes=1000):
        super(ResNet, self).__init__()
        self.inplanes = 64
        self.conv1 = Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)

        return x

def resnet50(**kwargs):
    """Constructs a ResNet-50 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
    return model

那么返回的resnet50就是用了MABN的resnet50。