來(lái)源丨手寫AI 

并行多分支結(jié)構(gòu)提取的特征具有更強(qiáng)的表征性；

具體可以回憶參考DenseNet、VOVNet、Res2Net以及PeleeNet（后續(xù)均會(huì)有解讀）。

并行多分支結(jié)構(gòu)會(huì)帶來(lái)更大別的計(jì)算開(kāi)銷；

具體可以參考CSPNet對(duì)此的分析。

使用 1×3 conv + 3×1 conv + 3×3 conv 代替原本一個(gè)的 3×3 conv的ACNet重參方法是有效的；

具體可以參考ACNet的分析。

答：有，DBB可以說(shuō)就是ACNet v2，全面升級(jí)！

Diverse Branch Block是繼ACNet的又一次對(duì)網(wǎng)絡(luò)結(jié)構(gòu)重參數(shù)化的探索，即ACNet v2，DBB設(shè)計(jì)了一個(gè)類似Inception的模塊，以多分支的結(jié)構(gòu)豐富卷積塊的特征空間，各分支結(jié)構(gòu)包括平均池化，多尺度卷積等。最后在推理階段前，把多分支結(jié)構(gòu)中進(jìn)行重參數(shù)化，融合成一個(gè)主分支。加快推理速度的同時(shí)，順帶提升一下精度！上圖給出了設(shè)計(jì)的DBB結(jié)構(gòu)示意圖。類似Inception,它采用1×1、1×1?K×K、1×1?AVG等組合方式對(duì)原始K×K卷積進(jìn)行增強(qiáng)。對(duì)于1×1?K×K分支，設(shè)置中間通道數(shù)等于輸入通道數(shù)并將1×1卷積初始化為Identity矩陣；其他分支則采用常規(guī)方式初始化。此外，在每個(gè)卷積后都添加BN層用于提供訓(xùn)練時(shí)的非線性，這對(duì)于性能提升很有必要。

以下是 DBB 的Pytorch實(shí)現(xiàn)：

import torch
import torch.nn as nn
import torch.nn.functional as F
from dbb_transforms import *


def conv_bn(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1,
            padding_mode='zeros'):
    conv_layer = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
                           stride=stride, padding=padding, dilation=dilation, groups=groups,
                           bias=False, padding_mode=padding_mode)
    bn_layer = nn.BatchNorm2d(num_features=out_channels, affine=True)
    se = nn.Sequential()
    se.add_module('conv', conv_layer)
    se.add_module('bn', bn_layer)
    return se


class IdentityBasedConv1x1(nn.Conv2d):
    def __init__(self, channels, groups=1):
        super(IdentityBasedConv1x1, self).__init__(in_channels=channels,
                                                   out_channels=channels,
                                                   kernel_size=1,
                                                   stride=1,
                                                   padding=0,
                                                   groups=groups,
                                                   bias=False)

        assert channels % groups == 0
        input_dim = channels // groups
        id_value = np.zeros((channels, input_dim, 1, 1))
        for i in range(channels):
            id_value[i, i % input_dim, 0, 0] = 1
        self.id_tensor = torch.from_numpy(id_value).type_as(self.weight)
        nn.init.zeros_(self.weight)

    def forward(self, input):
        kernel = self.weight + self.id_tensor.to(self.weight.device)
        result = F.conv2d(input,
                          kernel,
                          None,
                          stride=1,
                          padding=0,
                          dilation=self.dilation,
                          groups=self.groups)
        return result

    def get_actual_kernel(self):
        return self.weight + self.id_tensor.to(self.weight.device)


class BNAndPadLayer(nn.Module):
    def __init__(self,
                 pad_pixels,
                 num_features,
                 eps=1e-5,
                 momentum=0.1,
                 affine=True,
                 track_running_stats=True):
        super(BNAndPadLayer, self).__init__()
        self.bn = nn.BatchNorm2d(num_features,
                                 eps,
                                 momentum,
                                 affine,
                                 track_running_stats)
        self.pad_pixels = pad_pixels

    def forward(self, input):
        output = self.bn(input)
        if self.pad_pixels > 0:
            if self.bn.affine:
                pad_values = self.bn.bias.detach() - self.bn.running_mean * self.bn.weight.detach() / torch.sqrt(
                    self.bn.running_var + self.bn.eps)
            else:
                pad_values = - self.bn.running_mean / torch.sqrt(self.bn.running_var + self.bn.eps)
            output = F.pad(output, [self.pad_pixels] * 4)
            pad_values = pad_values.view(1, -1, 1, 1)
            output[:, :, 0:self.pad_pixels, :] = pad_values
            output[:, :, -self.pad_pixels:, :] = pad_values
            output[:, :, :, 0:self.pad_pixels] = pad_values
            output[:, :, :, -self.pad_pixels:] = pad_values
        return output

    @property
    def weight(self):
        return self.bn.weight

    @property
    def bias(self):
        return self.bn.bias

    @property
    def running_mean(self):
        return self.bn.running_mean

    @property
    def running_var(self):
        return self.bn.running_var

    @property
    def eps(self):
        return self.bn.eps


class DiverseBranchBlock(nn.Module):

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 stride=1,
                 padding=0,
                 dilation=1,
                 groups=1,
                 internal_channels_1x1_3x3=None,
                 deploy=False,
                 nonlinear=None,
                 single_init=False):
        super(DiverseBranchBlock, self).__init__()
        self.deploy = deploy

        if nonlinear is None:
            self.nonlinear = nn.Identity()
        else:
            self.nonlinear = nonlinear

        self.kernel_size = kernel_size
        self.out_channels = out_channels
        self.groups = groups
        assert padding == kernel_size // 2

        if deploy:
            self.dbb_reparam = nn.Conv2d(in_channels=in_channels,
                                         out_channels=out_channels,
                                         kernel_size=kernel_size,
                                         stride=stride,
                                         padding=padding,
                                         dilation=dilation,
                                         groups=groups,
                                         bias=True)

        else:
            self.dbb_origin = conv_bn(in_channels=in_channels,
                                      out_channels=out_channels,
                                      kernel_size=kernel_size,
                                      stride=stride,
                                      padding=padding,
                                      dilation=dilation,
                                      groups=groups)

            self.dbb_avg = nn.Sequential()
            if groups < out_channels:
                self.dbb_avg.add_module('conv',
                                        nn.Conv2d(in_channels=in_channels,
                                                  out_channels=out_channels,
                                                  kernel_size=1,
                                                  stride=1,
                                                  padding=0,
                                                  groups=groups,
                                                  bias=False))

                self.dbb_avg.add_module('bn',
                                        BNAndPadLayer(pad_pixels=padding,
                                                      num_features=out_channels))

                self.dbb_avg.add_module('avg',
                                        nn.AvgPool2d(kernel_size=kernel_size,
                                                     stride=stride,
                                                     padding=0))

                self.dbb_1x1 = conv_bn(in_channels=in_channels,
                                       out_channels=out_channels,
                                       kernel_size=1,
                                       stride=stride,
                                       padding=0,
                                       groups=groups)
            else:
                self.dbb_avg.add_module('avg',
                                        nn.AvgPool2d(kernel_size=kernel_size,
                                                     stride=stride,
                                                     padding=padding))

            self.dbb_avg.add_module('avgbn',
                                    nn.BatchNorm2d(out_channels))

            if internal_channels_1x1_3x3 is None:
                # For mobilenet, it is better to have 2X internal channels
                internal_channels_1x1_3x3 = in_channels if groups < out_channels else 2 * in_channels

            self.dbb_1x1_kxk = nn.Sequential()
            if internal_channels_1x1_3x3 == in_channels:
                self.dbb_1x1_kxk.add_module('idconv1',
                                            IdentityBasedConv1x1(channels=in_channels, groups=groups))
            else:
                self.dbb_1x1_kxk.add_module('conv1',
                                            nn.Conv2d(in_channels=in_channels,
           out_channels=internal_channels_1x1_3x3,
                                            kernel_size=1,
                                            stride=1,
                                            padding=0,
                                            groups=groups,
                                            bias=False))
            self.dbb_1x1_kxk.add_module('bn1',
                                        BNAndPadLayer(pad_pixels=padding,
               num_features=internal_channels_1x1_3x3,affine=True))
            self.dbb_1x1_kxk.add_module('conv2',
                                        nn.Conv2d(in_channels=internal_channels_1x1_3x3, 
              out_channels=out_channels,
                                                  kernel_size=kernel_size,
                                                  stride=stride,
                                                  padding=0,
                                                  groups=groups,
                                                  bias=False))
            self.dbb_1x1_kxk.add_module('bn2', nn.BatchNorm2d(out_channels))

        #   The experiments reported in the paper used the default initialization of bn.weight (all as 1).
        #   But changing the initialization may be useful in some cases.
        if single_init:
            #   Initialize the bn.weight of dbb_origin as 1 and others as 0.
            #   This is not the default setting.
            self.single_init()

    def get_equivalent_kernel_bias(self):
        k_origin, b_origin = transI_fusebn(self.dbb_origin.conv.weight,
                                           self.dbb_origin.bn)

        if hasattr(self, 'dbb_1x1'):
            # 按照方式1進(jìn)行conv+bn的融合
            k_1x1, b_1x1 = transI_fusebn(self.dbb_1x1.conv.weight,
                                         self.dbb_1x1.bn)
            # 按照方式方式6進(jìn)行多尺度卷積的合并
            k_1x1 = transVI_multiscale(k_1x1,
                                       self.kernel_size)
        else:
            k_1x1, b_1x1 = 0, 0

        if hasattr(self.dbb_1x1_kxk, 'idconv1'):
            k_1x1_kxk_first = self.dbb_1x1_kxk.idconv1.get_actual_kernel()
        else:
            k_1x1_kxk_first = self.dbb_1x1_kxk.conv1.weight
        # 按照方式1進(jìn)行conv+bn的融合
        k_1x1_kxk_first, b_1x1_kxk_first = transI_fusebn(k_1x1_kxk_first,
                                                         self.dbb_1x1_kxk.bn1)
        # 按照方式1進(jìn)行conv+bn的融合
        k_1x1_kxk_second, b_1x1_kxk_second = transI_fusebn(self.dbb_1x1_kxk.conv2.weight,
                                                           self.dbb_1x1_kxk.bn2)
        # 按照方式3進(jìn)行1x1卷積與kxk卷積的合并
        k_1x1_kxk_merged, b_1x1_kxk_merged = transIII_1x1_kxk(k_1x1_kxk_first,
                                                              b_1x1_kxk_first,
                                                              k_1x1_kxk_second,
                                                              b_1x1_kxk_second,
                                                              groups=self.groups)

        k_avg = transV_avg(self.out_channels, self.kernel_size, self.groups)
        # 按照方式1進(jìn)行conv+bn的融合
        k_1x1_avg_second, b_1x1_avg_second = transI_fusebn(k_avg.to(self.dbb_avg.avgbn.weight.device),
                                                           self.dbb_avg.avgbn)
        if hasattr(self.dbb_avg, 'conv'):
            # 按照方式1進(jìn)行conv+bn的融合
            k_1x1_avg_first, b_1x1_avg_first = transI_fusebn(self.dbb_avg.conv.weight,
                                                             self.dbb_avg.bn)
            # 按照方式3進(jìn)行1x1卷積與kxk卷積的合并
            k_1x1_avg_merged, b_1x1_avg_merged = transIII_1x1_kxk(k_1x1_avg_first,
                                                                  b_1x1_avg_first,
                                                                  k_1x1_avg_second,
                                                                  b_1x1_avg_second,
                                                                  groups=self.groups)
        else:
            k_1x1_avg_merged, b_1x1_avg_merged = k_1x1_avg_second, b_1x1_avg_second
        # 按照方式2進(jìn)行分支的合并
        return transII_addbranch((k_origin,
                                  k_1x1,
                                  k_1x1_kxk_merged,
                                  k_1x1_avg_merged),
                                 (b_origin,
                                  b_1x1,
                                  b_1x1_kxk_merged,
                                  b_1x1_avg_merged))

    def switch_to_deploy(self):
        if hasattr(self, 'dbb_reparam'):
            return
        kernel, bias = self.get_equivalent_kernel_bias()
        self.dbb_reparam = nn.Conv2d(in_channels=self.dbb_origin.conv.in_channels,
                                     out_channels=self.dbb_origin.conv.out_channels,
                                     kernel_size=self.dbb_origin.conv.kernel_size,
                                     stride=self.dbb_origin.conv.stride,
                                     padding=self.dbb_origin.conv.padding,
                                     dilation=self.dbb_origin.conv.dilation,
                                     groups=self.dbb_origin.conv.groups, bias=True)
        self.dbb_reparam.weight.data = kernel
        self.dbb_reparam.bias.data = bias
        for para in self.parameters():
            para.detach_()
        self.__delattr__('dbb_origin')
        self.__delattr__('dbb_avg')
        if hasattr(self, 'dbb_1x1'):
            self.__delattr__('dbb_1x1')
        self.__delattr__('dbb_1x1_kxk')

    def forward(self, inputs):

        if hasattr(self, 'dbb_reparam'):
            return self.nonlinear(self.dbb_reparam(inputs))

        out = self.dbb_origin(inputs)
        if hasattr(self, 'dbb_1x1'):
            out += self.dbb_1x1(inputs)
        out += self.dbb_avg(inputs)
        out += self.dbb_1x1_kxk(inputs)
        return self.nonlinear(out)

    def init_gamma(self, gamma_value):
        if hasattr(self, "dbb_origin"):
            torch.nn.init.constant_(self.dbb_origin.bn.weight,
                                    gamma_value)
        if hasattr(self, "dbb_1x1"):
            torch.nn.init.constant_(self.dbb_1x1.bn.weight,
                                    gamma_value)
        if hasattr(self, "dbb_avg"):
            torch.nn.init.constant_(self.dbb_avg.avgbn.weight,
                                    gamma_value)
        if hasattr(self, "dbb_1x1_kxk"):
            torch.nn.init.constant_(self.dbb_1x1_kxk.bn2.weight,
                                    gamma_value)

    def single_init(self):
        self.init_gamma(0.0)
        if hasattr(self, "dbb_origin"):
            torch.nn.init.constant_(self.dbb_origin.bn.weight, 1.0)

話不多說(shuō)，直接對(duì)比ONNX的輸出，就問(wèn)你香不香?。?！

[1].https://github.com/DingXiaoH/DiverseBranchBlock/blob/main/diversebranchblock.py

本文僅做學(xué)術(shù)分享，如有侵權(quán)，請(qǐng)聯(lián)系刪文。