第J6周：ResNeXt-50实战解析

weixin_46559155

于 2024-03-01 22:19:19 发布

阅读量849

点赞数 29

文章标签： python

本文链接：https://blog.csdn.net/weixin_46559155/article/details/136406987

版权

🍨 本文为🔗365天深度学习训练营中的学习记录博客
🍖 原作者：K同学啊|接辅导、项目定制

ResNeXt是有何凯明团队在2017年CVPR会议上提出来的新型图像分类网络。它是ResNet的升级版，在ResNet的基础上，引入了cardinality的概念，类似于ResNet。ResNeXt也有ResNeXt-50，ResNeXt-101版本。

1 模型结构

在ResNeXt的论文中，作者提出了当时普遍存在的一个问题，如果要提高模型的准确率，往往采取加深网络或者加宽网络的方法。虽然这种方法有效，但随之而来的，是网络设计的难度和计算开销的增加。为了一点精度的提升往往需要付出更大的代价，因此需要一个更好的策略，在不额外增加计算代价的情况下，提升网络的精度。由此，何等人提出了cardinality的概念。

下面是ResNet（左）与ResNeXt（右）block的差异。在ResNet中，输入的具有256个通道的特征经过1x1卷积压缩4倍到64个通道，之后3x3的卷积核用于处理特征，经1x1卷积扩大通道数与原特征残差连接后输出。ResNeXt也是相同的处理策略，但在ResNeXt中，输入的具有256个通道的特征被分为32个组，每组被压缩64倍到4个通道后进行处理。32个组相加后与原特征残差连接后输出。这里cardinatity指的是一个block中所具有的相同分支的数目。

2 分组卷积

ResNeXt中采用的分组卷积简单来说是将特征图分为不同的组，再对每组特征图分别进行卷积，这个操作可以有效的降低计算量。

在分组卷积中，每个卷积核只处理部分通道，比如下图中，红色卷积核只处理红色的通道，绿色卷积核只处理绿色的通道，黄色卷积核只处理黄色通道。此时每个卷积核有2个通道，每个卷积核生成一张特征图。

分组卷积的优势在于其参数开销，图3是其对比效果。

3 代码实现

3.1 分组卷积模块

class GroupedConvBlock(nn.Module):
    def __init__(self, in_channel, kernel_size=3, stride=1, groups=32):
        super(GroupedConvBlock, self).__init__()
        self.g_channel = in_channel//groups
        self.groups = groups
        self.conv = nn.Conv2d(self.g_channel, self.g_channel, kernel_size=3, stride=stride, padding=1, bias=False)
        self.norm = nn.BatchNorm2d(in_channel)
        self.relu = nn.ReLU(inplace=True)
        
    
    def forward(self, x):
        g_list = []
        # 分组进行卷积
        for c in range(self.groups):
            g = x[:,c*self.g_channel:(c+1)*self.g_channel,:,:]
            g = self.conv(g)
            g_list.append(g)
        x = torch.cat(g_list, dim=1)
        x = self.norm(x)
        x = self.relu(x)
        return x

3.2 定义残差单元

''' Residual Block '''
class Block(nn.Module):
    def __init__(self, in_channel, filters, kernel_size=3, stride=1, groups=32, conv_shortcut=True):
        super(Block, self).__init__()
        self.shortcut = conv_shortcut
        if self.shortcut:
            self.short = nn.Conv2d(in_channel, 2*filters, kernel_size=1, stride=stride, padding=0, bias=False)
        elif stride>1:
            self.short = nn.MaxPool2d(kernel_size=1, stride=stride, padding=0)
        else:
            self.short = nn.Identity()
        
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channel, filters, kernel_size=1, stride=1, bias=False),
            nn.BatchNorm2d(filters),
            nn.ReLU(True)
        )
        self.conv2 = GroupedConvBlock(in_channel=filters, kernel_size=kernel_size, stride=stride, groups=groups)
        self.conv3 = nn.Sequential(
            nn.Conv2d(filters, 2*filters, kernel_size=1, stride=1, bias=False),
            nn.BatchNorm2d(2*filters)
        )
        self.relu = nn.ReLU(inplace=True)
    
    def forward(self, x):
        if self.shortcut:
            x2 = self.short(x)
        else:
            x2 = self.short(x)
        x1 = self.conv1(x)
        x1 = self.conv2(x1)
        x1 = self.conv3(x1)
        x = x1 + x2
        x = self.relu(x)
        return x

3.3.堆叠残差单元

每个stack的第一个block的输入和输出的shape是不一致的，所以残差连接都需要使用1*1卷积升维后才能进行Add操作。

而其他block的输入和输出的shape是一致的，所以可以直接执行Add操作。

class Stack(nn.Module):
    def __init__(self, in_channel, filters, blocks, stride=2, groups=32):
        super(Stack, self).__init__()
        self.conv = nn.Sequential()
        self.conv.add_module(str(0), Block(in_channel, filters, stride=stride, groups=groups, conv_shortcut=True))
        for i in range(1, blocks):
            self.conv.add_module(str(i), Block(2*filters, filters, stride=1, groups=groups, conv_shortcut=False))
    
    def forward(self, x):
        x = self.conv(x)
        return x

3.4搭建ResNext-50网络

''' ResNeXt50 '''
class ResNeXt50(nn.Module):
    def __init__(self,
                 include_top=True,  # 是否包含位于网络顶部的全链接层
                 preact=False,  # 是否使用预激活
                 use_bias=True,  # 是否对卷积层使用偏置
                 input_shape=[32, 3, 224, 224],
                 classes=1000,
                 pooling=None):  # 用于分类图像的可选类数
        super(ResNeXt50, self).__init__()
        
        self.conv1 = nn.Sequential()
        self.conv1.add_module('conv', nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=use_bias, padding_mode='zeros'))
        if not preact:
            self.conv1.add_module('bn', nn.BatchNorm2d(64))
            self.conv1.add_module('relu', nn.ReLU())
        self.conv1.add_module('max_pool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
        
        self.conv2 = Stack(64, 128, 3, stride=1)
        self.conv3 = Stack(256, 256, 4, stride=2)
        self.conv4 = Stack(512, 512, 6, stride=2)
        self.conv5 = Stack(1024, 1024, 3, stride=2)
        
        self.post = nn.Sequential()
        if preact:
            self.post.add_module('bn', nn.BatchNorm2d(2048))
            self.post.add_module('relu', nn.ReLU())
        if include_top:
            self.post.add_module('avg_pool', nn.AdaptiveAvgPool2d((1, 1)))
            self.post.add_module('flatten', nn.Flatten())
            self.post.add_module('fc', nn.Linear(2048, classes))
        else:
            if pooling=='avg':
                self.post.add_module('avg_pool', nn.AdaptiveAvgPool2d((1, 1)))
            elif pooling=='max':
                self.post.add_module('max_pool', nn.AdaptiveMaxPool2d((1, 1)))
    
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        x = self.conv5(x)
        x = self.post(x)
        return x

3.5查看模型摘要

''' 调用并将模型转移到GPU中（我们模型运行均在GPU中进行） '''
model = ResNeXt50(n_class=num_classes).to(device)
#model = ResNeXt50(n_class=num_classes).to(device)
''' 显示网络结构 '''
torchsummary.summary(model, (32, 3, 224, 224))
#torchinfo.summary(model)
print(model)

----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,472
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
MaxPool2d-4 [-1, 64, 56, 56] 0
Conv2d-5 [-1, 256, 56, 56] 16,384
Conv2d-6 [-1, 128, 56, 56] 8,192
BatchNorm2d-7 [-1, 128, 56, 56] 256
ReLU-8 [-1, 128, 56, 56] 0
Conv2d-9 [-1, 4, 56, 56] 144
Conv2d-10 [-1, 4, 56, 56] 144
Conv2d-11 [-1, 4, 56, 56] 144
Conv2d-12 [-1, 4, 56, 56] 144
Conv2d-13 [-1, 4, 56, 56] 144
Conv2d-14 [-1, 4, 56, 56] 144
Conv2d-15 [-1, 4, 56, 56] 144
Conv2d-16 [-1, 4, 56, 56] 144
Conv2d-17 [-1, 4, 56, 56] 144
Conv2d-18 [-1, 4, 56, 56] 144
Conv2d-19 [-1, 4, 56, 56] 144
Conv2d-20 [-1, 4, 56, 56] 144
Conv2d-21 [-1, 4, 56, 56] 144
Conv2d-22 [-1, 4, 56, 56] 144
Conv2d-23 [-1, 4, 56, 56] 144
Conv2d-24 [-1, 4, 56, 56] 144
Conv2d-25 [-1, 4, 56, 56] 144
Conv2d-26 [-1, 4, 56, 56] 144
Conv2d-27 [-1, 4, 56, 56] 144
Conv2d-28 [-1, 4, 56, 56] 144
Conv2d-29 [-1, 4, 56, 56] 144
Conv2d-30 [-1, 4, 56, 56] 144
Conv2d-31 [-1, 4, 56, 56] 144
Conv2d-32 [-1, 4, 56, 56] 144
Conv2d-33 [-1, 4, 56, 56] 144
Conv2d-34 [-1, 4, 56, 56] 144
Conv2d-35 [-1, 4, 56, 56] 144
Conv2d-36 [-1, 4, 56, 56] 144
Conv2d-37 [-1, 4, 56, 56] 144
Conv2d-38 [-1, 4, 56, 56] 144
Conv2d-39 [-1, 4, 56, 56] 144
Conv2d-40 [-1, 4, 56, 56] 144
BatchNorm2d-41 [-1, 128, 56, 56] 256
ReLU-42 [-1, 128, 56, 56] 0
GroupedConvBlock-43 [-1, 128, 56, 56] 0
Conv2d-44 [-1, 256, 56, 56] 32,768
BatchNorm2d-45 [-1, 256, 56, 56] 512
ReLU-46 [-1, 256, 56, 56] 0
Block-47 [-1, 256, 56, 56] 0
Identity-48 [-1, 256, 56, 56] 0
Conv2d-49 [-1, 128, 56, 56] 32,768
BatchNorm2d-50 [-1, 128, 56, 56] 256
ReLU-51 [-1, 128, 56, 56] 0
Conv2d-52 [-1, 4, 56, 56] 144
Conv2d-53 [-1, 4, 56, 56] 144
Conv2d-54 [-1, 4, 56, 56] 144
Conv2d-55 [-1, 4, 56, 56] 144
Conv2d-56 [-1, 4, 56, 56] 144
Conv2d-57 [-1, 4, 56, 56] 144
Conv2d-58 [-1, 4, 56, 56] 144
Conv2d-59 [-1, 4, 56, 56] 144
Conv2d-60 [-1, 4, 56, 56] 144
Conv2d-61 [-1, 4, 56, 56] 144
Conv2d-62 [-1, 4, 56, 56] 144
Conv2d-63 [-1, 4, 56, 56] 144
Conv2d-64 [-1, 4, 56, 56] 144
Conv2d-65 [-1, 4, 56, 56] 144
Conv2d-66 [-1, 4, 56, 56] 144
Conv2d-67 [-1, 4, 56, 56] 144
Conv2d-68 [-1, 4, 56, 56] 144
Conv2d-69 [-1, 4, 56, 56] 144
Conv2d-70 [-1, 4, 56, 56] 144
Conv2d-71 [-1, 4, 56, 56] 144
Conv2d-72 [-1, 4, 56, 56] 144
Conv2d-73 [-1, 4, 56, 56] 144
Conv2d-74 [-1, 4, 56, 56] 144
Conv2d-75 [-1, 4, 56, 56] 144
Conv2d-76 [-1, 4, 56, 56] 144
Conv2d-77 [-1, 4, 56, 56] 144
Conv2d-78 [-1, 4, 56, 56] 144
Conv2d-79 [-1, 4, 56, 56] 144
Conv2d-80 [-1, 4, 56, 56] 144
Conv2d-81 [-1, 4, 56, 56] 144
Conv2d-82 [-1, 4, 56, 56] 144
Conv2d-83 [-1, 4, 56, 56] 144
BatchNorm2d-84 [-1, 128, 56, 56] 256
ReLU-85 [-1, 128, 56, 56] 0
GroupedConvBlock-86 [-1, 128, 56, 56] 0
Conv2d-87 [-1, 256, 56, 56] 32,768
BatchNorm2d-88 [-1, 256, 56, 56] 512
ReLU-89 [-1, 256, 56, 56] 0
Block-90 [-1, 256, 56, 56] 0
Identity-91 [-1, 256, 56, 56] 0
Conv2d-92 [-1, 128, 56, 56] 32,768
BatchNorm2d-93 [-1, 128, 56, 56] 256
ReLU-94 [-1, 128, 56, 56] 0
Conv2d-95 [-1, 4, 56, 56] 144
Conv2d-96 [-1, 4, 56, 56] 144
Conv2d-97 [-1, 4, 56, 56] 144
Conv2d-98 [-1, 4, 56, 56] 144
Conv2d-99 [-1, 4, 56, 56] 144
Conv2d-100 [-1, 4, 56, 56] 144
Conv2d-101 [-1, 4, 56, 56] 144
Conv2d-102 [-1, 4, 56, 56] 144
Conv2d-103 [-1, 4, 56, 56] 144
Conv2d-104 [-1, 4, 56, 56] 144
Conv2d-105 [-1, 4, 56, 56] 144
Conv2d-106 [-1, 4, 56, 56] 144
Conv2d-107 [-1, 4, 56, 56] 144
Conv2d-108 [-1, 4, 56, 56] 144
Conv2d-109 [-1, 4, 56, 56] 144
Conv2d-110 [-1, 4, 56, 56] 144
Conv2d-111 [-1, 4, 56, 56] 144
Conv2d-112 [-1, 4, 56, 56] 144
Conv2d-113 [-1, 4, 56, 56] 144
Conv2d-114 [-1, 4, 56, 56] 144
Conv2d-115 [-1, 4, 56, 56] 144
Conv2d-116 [-1, 4, 56, 56] 144
Conv2d-117 [-1, 4, 56, 56] 144
Conv2d-118 [-1, 4, 56, 56] 144
Conv2d-119 [-1, 4, 56, 56] 144
Conv2d-120 [-1, 4, 56, 56] 144
Conv2d-121 [-1, 4, 56, 56] 144
Conv2d-122 [-1, 4, 56, 56] 144
Conv2d-123 [-1, 4, 56, 56] 144
Conv2d-124 [-1, 4, 56, 56] 144
Conv2d-125 [-1, 4, 56, 56] 144
Conv2d-126 [-1, 4, 56, 56] 144
BatchNorm2d-127 [-1, 128, 56, 56] 256
ReLU-128 [-1, 128, 56, 56] 0
GroupedConvBlock-129 [-1, 128, 56, 56] 0
Conv2d-130 [-1, 256, 56, 56] 32,768
BatchNorm2d-131 [-1, 256, 56, 56] 512
ReLU-132 [-1, 256, 56, 56] 0
Block-133 [-1, 256, 56, 56] 0
Stack-134 [-1, 256, 56, 56] 0
Conv2d-135 [-1, 512, 28, 28] 131,072
Conv2d-136 [-1, 256, 56, 56] 65,536
BatchNorm2d-137 [-1, 256, 56, 56] 512
ReLU-138 [-1, 256, 56, 56] 0
Conv2d-139 [-1, 8, 28, 28] 576
Conv2d-140 [-1, 8, 28, 28] 576
Conv2d-141 [-1, 8, 28, 28] 576
Conv2d-142 [-1, 8, 28, 28] 576
Conv2d-143 [-1, 8, 28, 28] 576
Conv2d-144 [-1, 8, 28, 28] 576
Conv2d-145 [-1, 8, 28, 28] 576
Conv2d-146 [-1, 8, 28, 28] 576
Conv2d-147 [-1, 8, 28, 28] 576
Conv2d-148 [-1, 8, 28, 28] 576
Conv2d-149 [-1, 8, 28, 28] 576
Conv2d-150 [-1, 8, 28, 28] 576
Conv2d-151 [-1, 8, 28, 28] 576
Conv2d-152 [-1, 8, 28, 28] 576
Conv2d-153 [-1, 8, 28, 28] 576
Conv2d-154 [-1, 8, 28, 28] 576
Conv2d-155 [-1, 8, 28, 28] 576
Conv2d-156 [-1, 8, 28, 28] 576
Conv2d-157 [-1, 8, 28, 28] 576
Conv2d-158 [-1, 8, 28, 28] 576
Conv2d-159 [-1, 8, 28, 28] 576
Conv2d-160 [-1, 8, 28, 28] 576
Conv2d-161 [-1, 8, 28, 28] 576
Conv2d-162 [-1, 8, 28, 28] 576
Conv2d-163 [-1, 8, 28, 28] 576
Conv2d-164 [-1, 8, 28, 28] 576
Conv2d-165 [-1, 8, 28, 28] 576
Conv2d-166 [-1, 8, 28, 28] 576
Conv2d-167 [-1, 8, 28, 28] 576
Conv2d-168 [-1, 8, 28, 28] 576
Conv2d-169 [-1, 8, 28, 28] 576
Conv2d-170 [-1, 8, 28, 28] 576
BatchNorm2d-171 [-1, 256, 28, 28] 512
ReLU-172 [-1, 256, 28, 28] 0
GroupedConvBlock-173 [-1, 256, 28, 28] 0
Conv2d-174 [-1, 512, 28, 28] 131,072
BatchNorm2d-175 [-1, 512, 28, 28] 1,024
ReLU-176 [-1, 512, 28, 28] 0
Block-177 [-1, 512, 28, 28] 0
Identity-178 [-1, 512, 28, 28] 0
Conv2d-179 [-1, 256, 28, 28] 131,072
BatchNorm2d-180 [-1, 256, 28, 28] 512
ReLU-181 [-1, 256, 28, 28] 0
Conv2d-182 [-1, 8, 28, 28] 576
Conv2d-183 [-1, 8, 28, 28] 576
Conv2d-184 [-1, 8, 28, 28] 576
Conv2d-185 [-1, 8, 28, 28] 576
Conv2d-186 [-1, 8, 28, 28] 576
Conv2d-187 [-1, 8, 28, 28] 576
Conv2d-188 [-1, 8, 28, 28] 576
Conv2d-189 [-1, 8, 28, 28] 576
Conv2d-190 [-1, 8, 28, 28] 576
Conv2d-191 [-1, 8, 28, 28] 576
Conv2d-192 [-1, 8, 28, 28] 576
Conv2d-193 [-1, 8, 28, 28] 576
Conv2d-194 [-1, 8, 28, 28] 576
Conv2d-195 [-1, 8, 28, 28] 576
Conv2d-196 [-1, 8, 28, 28] 576
Conv2d-197 [-1, 8, 28, 28] 576
Conv2d-198 [-1, 8, 28, 28] 576
Conv2d-199 [-1, 8, 28, 28] 576
Conv2d-200 [-1, 8, 28, 28] 576
Conv2d-201 [-1, 8, 28, 28] 576
Conv2d-202 [-1, 8, 28, 28] 576
Conv2d-203 [-1, 8, 28, 28] 576
Conv2d-204 [-1, 8, 28, 28] 576
Conv2d-205 [-1, 8, 28, 28] 576
Conv2d-206 [-1, 8, 28, 28] 576
Conv2d-207 [-1, 8, 28, 28] 576
Conv2d-208 [-1, 8, 28, 28] 576
Conv2d-209 [-1, 8, 28, 28] 576
Conv2d-210 [-1, 8, 28, 28] 576
Conv2d-211 [-1, 8, 28, 28] 576
Conv2d-212 [-1, 8, 28, 28] 576
Conv2d-213 [-1, 8, 28, 28] 576
BatchNorm2d-214 [-1, 256, 28, 28] 512
ReLU-215 [-1, 256, 28, 28] 0
GroupedConvBlock-216 [-1, 256, 28, 28] 0
Conv2d-217 [-1, 512, 28, 28] 131,072
BatchNorm2d-218 [-1, 512, 28, 28] 1,024
ReLU-219 [-1, 512, 28, 28] 0
Block-220 [-1, 512, 28, 28] 0
Identity-221 [-1, 512, 28, 28] 0
Conv2d-222 [-1, 256, 28, 28] 131,072
BatchNorm2d-223 [-1, 256, 28, 28] 512
ReLU-224 [-1, 256, 28, 28] 0
Conv2d-225 [-1, 8, 28, 28] 576
Conv2d-226 [-1, 8, 28, 28] 576
Conv2d-227 [-1, 8, 28, 28] 576
Conv2d-228 [-1, 8, 28, 28] 576
Conv2d-229 [-1, 8, 28, 28] 576
Conv2d-230 [-1, 8, 28, 28] 576
Conv2d-231 [-1, 8, 28, 28] 576
Conv2d-232 [-1, 8, 28, 28] 576
Conv2d-233 [-1, 8, 28, 28] 576
Conv2d-234 [-1, 8, 28, 28] 576
Conv2d-235 [-1, 8, 28, 28] 576
Conv2d-236 [-1, 8, 28, 28] 576
Conv2d-237 [-1, 8, 28, 28] 576
Conv2d-238 [-1, 8, 28, 28] 576
Conv2d-239 [-1, 8, 28, 28] 576
Conv2d-240 [-1, 8, 28, 28] 576
Conv2d-241 [-1, 8, 28, 28] 576
Conv2d-242 [-1, 8, 28, 28] 576
Conv2d-243 [-1, 8, 28, 28] 576
Conv2d-244 [-1, 8, 28, 28] 576
Conv2d-245 [-1, 8, 28, 28] 576
Conv2d-246 [-1, 8, 28, 28] 576
Conv2d-247 [-1, 8, 28, 28] 576
Conv2d-248 [-1, 8, 28, 28] 576
Conv2d-249 [-1, 8, 28, 28] 576
Conv2d-250 [-1, 8, 28, 28] 576
Conv2d-251 [-1, 8, 28, 28] 576
Conv2d-252 [-1, 8, 28, 28] 576
Conv2d-253 [-1, 8, 28, 28] 576
Conv2d-254 [-1, 8, 28, 28] 576
Conv2d-255 [-1, 8, 28, 28] 576
Conv2d-256 [-1, 8, 28, 28] 576
BatchNorm2d-257 [-1, 256, 28, 28] 512
ReLU-258 [-1, 256, 28, 28] 0
GroupedConvBlock-259 [-1, 256, 28, 28] 0
Conv2d-260 [-1, 512, 28, 28] 131,072
BatchNorm2d-261 [-1, 512, 28, 28] 1,024
ReLU-262 [-1, 512, 28, 28] 0
Block-263 [-1, 512, 28, 28] 0
Identity-264 [-1, 512, 28, 28] 0
Conv2d-265 [-1, 256, 28, 28] 131,072
BatchNorm2d-266 [-1, 256, 28, 28] 512
ReLU-267 [-1, 256, 28, 28] 0
Conv2d-268 [-1, 8, 28, 28] 576
Conv2d-269 [-1, 8, 28, 28] 576
Conv2d-270 [-1, 8, 28, 28] 576
Conv2d-271 [-1, 8, 28, 28] 576
Conv2d-272 [-1, 8, 28, 28] 576
Conv2d-273 [-1, 8, 28, 28] 576
Conv2d-274 [-1, 8, 28, 28] 576
Conv2d-275 [-1, 8, 28, 28] 576
Conv2d-276 [-1, 8, 28, 28] 576
Conv2d-277 [-1, 8, 28, 28] 576
Conv2d-278 [-1, 8, 28, 28] 576
Conv2d-279 [-1, 8, 28, 28] 576
Conv2d-280 [-1, 8, 28, 28] 576
Conv2d-281 [-1, 8, 28, 28] 576
Conv2d-282 [-1, 8, 28, 28] 576
Conv2d-283 [-1, 8, 28, 28] 576
Conv2d-284 [-1, 8, 28, 28] 576
Conv2d-285 [-1, 8, 28, 28] 576
Conv2d-286 [-1, 8, 28, 28] 576
Conv2d-287 [-1, 8, 28, 28] 576
Conv2d-288 [-1, 8, 28, 28] 576
Conv2d-289 [-1, 8, 28, 28] 576
Conv2d-290 [-1, 8, 28, 28] 576
Conv2d-291 [-1, 8, 28, 28] 576
Conv2d-292 [-1, 8, 28, 28] 576
Conv2d-293 [-1, 8, 28, 28] 576
Conv2d-294 [-1, 8, 28, 28] 576
Conv2d-295 [-1, 8, 28, 28] 576
Conv2d-296 [-1, 8, 28, 28] 576
Conv2d-297 [-1, 8, 28, 28] 576
Conv2d-298 [-1, 8, 28, 28] 576
Conv2d-299 [-1, 8, 28, 28] 576
BatchNorm2d-300 [-1, 256, 28, 28] 512
ReLU-301 [-1, 256, 28, 28] 0
GroupedConvBlock-302 [-1, 256, 28, 28] 0
Conv2d-303 [-1, 512, 28, 28] 131,072
BatchNorm2d-304 [-1, 512, 28, 28] 1,024
ReLU-305 [-1, 512, 28, 28] 0
Block-306 [-1, 512, 28, 28] 0
Stack-307 [-1, 512, 28, 28] 0
Conv2d-308 [-1, 1024, 14, 14] 524,288
Conv2d-309 [-1, 512, 28, 28] 262,144
BatchNorm2d-310 [-1, 512, 28, 28] 1,024
ReLU-311 [-1, 512, 28, 28] 0
Conv2d-312 [-1, 16, 14, 14] 2,304
Conv2d-313 [-1, 16, 14, 14] 2,304
Conv2d-314 [-1, 16, 14, 14] 2,304
Conv2d-315 [-1, 16, 14, 14] 2,304
Conv2d-316 [-1, 16, 14, 14] 2,304
Conv2d-317 [-1, 16, 14, 14] 2,304
Conv2d-318 [-1, 16, 14, 14] 2,304
Conv2d-319 [-1, 16, 14, 14] 2,304
Conv2d-320 [-1, 16, 14, 14] 2,304
Conv2d-321 [-1, 16, 14, 14] 2,304
Conv2d-322 [-1, 16, 14, 14] 2,304
Conv2d-323 [-1, 16, 14, 14] 2,304
Conv2d-324 [-1, 16, 14, 14] 2,304
Conv2d-325 [-1, 16, 14, 14] 2,304
Conv2d-326 [-1, 16, 14, 14] 2,304
Conv2d-327 [-1, 16, 14, 14] 2,304
Conv2d-328 [-1, 16, 14, 14] 2,304
Conv2d-329 [-1, 16, 14, 14] 2,304
Conv2d-330 [-1, 16, 14, 14] 2,304
Conv2d-331 [-1, 16, 14, 14] 2,304
Conv2d-332 [-1, 16, 14, 14] 2,304
Conv2d-333 [-1, 16, 14, 14] 2,304
Conv2d-334 [-1, 16, 14, 14] 2,304
Conv2d-335 [-1, 16, 14, 14] 2,304
Conv2d-336 [-1, 16, 14, 14] 2,304
Conv2d-337 [-1, 16, 14, 14] 2,304
Conv2d-338 [-1, 16, 14, 14] 2,304
Conv2d-339 [-1, 16, 14, 14] 2,304
Conv2d-340 [-1, 16, 14, 14] 2,304
Conv2d-341 [-1, 16, 14, 14] 2,304
Conv2d-342 [-1, 16, 14, 14] 2,304
Conv2d-343 [-1, 16, 14, 14] 2,304
BatchNorm2d-344 [-1, 512, 14, 14] 1,024
ReLU-345 [-1, 512, 14, 14] 0
GroupedConvBlock-346 [-1, 512, 14, 14] 0
Conv2d-347 [-1, 1024, 14, 14] 524,288
BatchNorm2d-348 [-1, 1024, 14, 14] 2,048
ReLU-349 [-1, 1024, 14, 14] 0
Block-350 [-1, 1024, 14, 14] 0
Identity-351 [-1, 1024, 14, 14] 0
Conv2d-352 [-1, 512, 14, 14] 524,288
BatchNorm2d-353 [-1, 512, 14, 14] 1,024
ReLU-354 [-1, 512, 14, 14] 0
Conv2d-355 [-1, 16, 14, 14] 2,304
Conv2d-356 [-1, 16, 14, 14] 2,304
Conv2d-357 [-1, 16, 14, 14] 2,304
Conv2d-358 [-1, 16, 14, 14] 2,304
Conv2d-359 [-1, 16, 14, 14] 2,304
Conv2d-360 [-1, 16, 14, 14] 2,304
Conv2d-361 [-1, 16, 14, 14] 2,304
Conv2d-362 [-1, 16, 14, 14] 2,304
Conv2d-363 [-1, 16, 14, 14] 2,304
Conv2d-364 [-1, 16, 14, 14] 2,304
Conv2d-365 [-1, 16, 14, 14] 2,304
Conv2d-366 [-1, 16, 14, 14] 2,304
Conv2d-367 [-1, 16, 14, 14] 2,304
Conv2d-368 [-1, 16, 14, 14] 2,304
Conv2d-369 [-1, 16, 14, 14] 2,304
Conv2d-370 [-1, 16, 14, 14] 2,304
Conv2d-371 [-1, 16, 14, 14] 2,304
Conv2d-372 [-1, 16, 14, 14] 2,304
Conv2d-373 [-1, 16, 14, 14] 2,304
Conv2d-374 [-1, 16, 14, 14] 2,304
Conv2d-375 [-1, 16, 14, 14] 2,304
Conv2d-376 [-1, 16, 14, 14] 2,304
Conv2d-377 [-1, 16, 14, 14] 2,304
Conv2d-378 [-1, 16, 14, 14] 2,304
Conv2d-379 [-1, 16, 14, 14] 2,304
Conv2d-380 [-1, 16, 14, 14] 2,304
Conv2d-381 [-1, 16, 14, 14] 2,304
Conv2d-382 [-1, 16, 14, 14] 2,304
Conv2d-383 [-1, 16, 14, 14] 2,304
Conv2d-384 [-1, 16, 14, 14] 2,304
Conv2d-385 [-1, 16, 14, 14] 2,304
Conv2d-386 [-1, 16, 14, 14] 2,304
BatchNorm2d-387 [-1, 512, 14, 14] 1,024
ReLU-388 [-1, 512, 14, 14] 0
GroupedConvBlock-389 [-1, 512, 14, 14] 0
Conv2d-390 [-1, 1024, 14, 14] 524,288
BatchNorm2d-391 [-1, 1024, 14, 14] 2,048
ReLU-392 [-1, 1024, 14, 14] 0
Block-393 [-1, 1024, 14, 14] 0
Identity-394 [-1, 1024, 14, 14] 0
Conv2d-395 [-1, 512, 14, 14] 524,288
BatchNorm2d-396 [-1, 512, 14, 14] 1,024
ReLU-397 [-1, 512, 14, 14] 0
Conv2d-398 [-1, 16, 14, 14] 2,304
Conv2d-399 [-1, 16, 14, 14] 2,304
Conv2d-400 [-1, 16, 14, 14] 2,304
Conv2d-401 [-1, 16, 14, 14] 2,304
Conv2d-402 [-1, 16, 14, 14] 2,304
Conv2d-403 [-1, 16, 14, 14] 2,304
Conv2d-404 [-1, 16, 14, 14] 2,304
Conv2d-405 [-1, 16, 14, 14] 2,304
Conv2d-406 [-1, 16, 14, 14] 2,304
Conv2d-407 [-1, 16, 14, 14] 2,304
Conv2d-408 [-1, 16, 14, 14] 2,304
Conv2d-409 [-1, 16, 14, 14] 2,304
Conv2d-410 [-1, 16, 14, 14] 2,304
Conv2d-411 [-1, 16, 14, 14] 2,304
Conv2d-412 [-1, 16, 14, 14] 2,304
Conv2d-413 [-1, 16, 14, 14] 2,304
Conv2d-414 [-1, 16, 14, 14] 2,304
Conv2d-415 [-1, 16, 14, 14] 2,304
Conv2d-416 [-1, 16, 14, 14] 2,304
Conv2d-417 [-1, 16, 14, 14] 2,304
Conv2d-418 [-1, 16, 14, 14] 2,304
Conv2d-419 [-1, 16, 14, 14] 2,304
Conv2d-420 [-1, 16, 14, 14] 2,304
Conv2d-421 [-1, 16, 14, 14] 2,304
Conv2d-422 [-1, 16, 14, 14] 2,304
Conv2d-423 [-1, 16, 14, 14] 2,304
Conv2d-424 [-1, 16, 14, 14] 2,304
Conv2d-425 [-1, 16, 14, 14] 2,304
Conv2d-426 [-1, 16, 14, 14] 2,304
Conv2d-427 [-1, 16, 14, 14] 2,304
Conv2d-428 [-1, 16, 14, 14] 2,304
Conv2d-429 [-1, 16, 14, 14] 2,304
BatchNorm2d-430 [-1, 512, 14, 14] 1,024
ReLU-431 [-1, 512, 14, 14] 0
GroupedConvBlock-432 [-1, 512, 14, 14] 0
Conv2d-433 [-1, 1024, 14, 14] 524,288
BatchNorm2d-434 [-1, 1024, 14, 14] 2,048
ReLU-435 [-1, 1024, 14, 14] 0
Block-436 [-1, 1024, 14, 14] 0
Identity-437 [-1, 1024, 14, 14] 0
Conv2d-438 [-1, 512, 14, 14] 524,288
BatchNorm2d-439 [-1, 512, 14, 14] 1,024
ReLU-440 [-1, 512, 14, 14] 0
Conv2d-441 [-1, 16, 14, 14] 2,304
Conv2d-442 [-1, 16, 14, 14] 2,304
Conv2d-443 [-1, 16, 14, 14] 2,304
Conv2d-444 [-1, 16, 14, 14] 2,304
Conv2d-445 [-1, 16, 14, 14] 2,304
Conv2d-446 [-1, 16, 14, 14] 2,304
Conv2d-447 [-1, 16, 14, 14] 2,304
Conv2d-448 [-1, 16, 14, 14] 2,304
Conv2d-449 [-1, 16, 14, 14] 2,304
Conv2d-450 [-1, 16, 14, 14] 2,304
Conv2d-451 [-1, 16, 14, 14] 2,304
Conv2d-452 [-1, 16, 14, 14] 2,304
Conv2d-453 [-1, 16, 14, 14] 2,304
Conv2d-454 [-1, 16, 14, 14] 2,304
Conv2d-455 [-1, 16, 14, 14] 2,304
Conv2d-456 [-1, 16, 14, 14] 2,304
Conv2d-457 [-1, 16, 14, 14] 2,304
Conv2d-458 [-1, 16, 14, 14] 2,304
Conv2d-459 [-1, 16, 14, 14] 2,304
Conv2d-460 [-1, 16, 14, 14] 2,304
Conv2d-461 [-1, 16, 14, 14] 2,304
Conv2d-462 [-1, 16, 14, 14] 2,304
Conv2d-463 [-1, 16, 14, 14] 2,304
Conv2d-464 [-1, 16, 14, 14] 2,304
Conv2d-465 [-1, 16, 14, 14] 2,304
Conv2d-466 [-1, 16, 14, 14] 2,304
Conv2d-467 [-1, 16, 14, 14] 2,304
Conv2d-468 [-1, 16, 14, 14] 2,304
Conv2d-469 [-1, 16, 14, 14] 2,304
Conv2d-470 [-1, 16, 14, 14] 2,304
Conv2d-471 [-1, 16, 14, 14] 2,304
Conv2d-472 [-1, 16, 14, 14] 2,304
BatchNorm2d-473 [-1, 512, 14, 14] 1,024
ReLU-474 [-1, 512, 14, 14] 0
GroupedConvBlock-475 [-1, 512, 14, 14] 0
Conv2d-476 [-1, 1024, 14, 14] 524,288
BatchNorm2d-477 [-1, 1024, 14, 14] 2,048
ReLU-478 [-1, 1024, 14, 14] 0
Block-479 [-1, 1024, 14, 14] 0
Identity-480 [-1, 1024, 14, 14] 0
Conv2d-481 [-1, 512, 14, 14] 524,288
BatchNorm2d-482 [-1, 512, 14, 14] 1,024
ReLU-483 [-1, 512, 14, 14] 0
Conv2d-484 [-1, 16, 14, 14] 2,304
Conv2d-485 [-1, 16, 14, 14] 2,304
Conv2d-486 [-1, 16, 14, 14] 2,304
Conv2d-487 [-1, 16, 14, 14] 2,304
Conv2d-488 [-1, 16, 14, 14] 2,304
Conv2d-489 [-1, 16, 14, 14] 2,304
Conv2d-490 [-1, 16, 14, 14] 2,304
Conv2d-491 [-1, 16, 14, 14] 2,304
Conv2d-492 [-1, 16, 14, 14] 2,304
Conv2d-493 [-1, 16, 14, 14] 2,304
Conv2d-494 [-1, 16, 14, 14] 2,304
Conv2d-495 [-1, 16, 14, 14] 2,304
Conv2d-496 [-1, 16, 14, 14] 2,304
Conv2d-497 [-1, 16, 14, 14] 2,304
Conv2d-498 [-1, 16, 14, 14] 2,304
Conv2d-499 [-1, 16, 14, 14] 2,304
Conv2d-500 [-1, 16, 14, 14] 2,304
Conv2d-501 [-1, 16, 14, 14] 2,304
Conv2d-502 [-1, 16, 14, 14] 2,304
Conv2d-503 [-1, 16, 14, 14] 2,304
Conv2d-504 [-1, 16, 14, 14] 2,304
Conv2d-505 [-1, 16, 14, 14] 2,304
Conv2d-506 [-1, 16, 14, 14] 2,304
Conv2d-507 [-1, 16, 14, 14] 2,304
Conv2d-508 [-1, 16, 14, 14] 2,304
Conv2d-509 [-1, 16, 14, 14] 2,304
Conv2d-510 [-1, 16, 14, 14] 2,304
Conv2d-511 [-1, 16, 14, 14] 2,304
Conv2d-512 [-1, 16, 14, 14] 2,304
Conv2d-513 [-1, 16, 14, 14] 2,304
Conv2d-514 [-1, 16, 14, 14] 2,304
Conv2d-515 [-1, 16, 14, 14] 2,304
BatchNorm2d-516 [-1, 512, 14, 14] 1,024
ReLU-517 [-1, 512, 14, 14] 0
GroupedConvBlock-518 [-1, 512, 14, 14] 0
Conv2d-519 [-1, 1024, 14, 14] 524,288
BatchNorm2d-520 [-1, 1024, 14, 14] 2,048
ReLU-521 [-1, 1024, 14, 14] 0
Block-522 [-1, 1024, 14, 14] 0
Identity-523 [-1, 1024, 14, 14] 0
Conv2d-524 [-1, 512, 14, 14] 524,288
BatchNorm2d-525 [-1, 512, 14, 14] 1,024
ReLU-526 [-1, 512, 14, 14] 0
Conv2d-527 [-1, 16, 14, 14] 2,304
Conv2d-528 [-1, 16, 14, 14] 2,304
Conv2d-529 [-1, 16, 14, 14] 2,304
Conv2d-530 [-1, 16, 14, 14] 2,304
Conv2d-531 [-1, 16, 14, 14] 2,304
Conv2d-532 [-1, 16, 14, 14] 2,304
Conv2d-533 [-1, 16, 14, 14] 2,304
Conv2d-534 [-1, 16, 14, 14] 2,304
Conv2d-535 [-1, 16, 14, 14] 2,304
Conv2d-536 [-1, 16, 14, 14] 2,304
Conv2d-537 [-1, 16, 14, 14] 2,304
Conv2d-538 [-1, 16, 14, 14] 2,304
Conv2d-539 [-1, 16, 14, 14] 2,304
Conv2d-540 [-1, 16, 14, 14] 2,304
Conv2d-541 [-1, 16, 14, 14] 2,304
Conv2d-542 [-1, 16, 14, 14] 2,304
Conv2d-543 [-1, 16, 14, 14] 2,304
Conv2d-544 [-1, 16, 14, 14] 2,304
Conv2d-545 [-1, 16, 14, 14] 2,304
Conv2d-546 [-1, 16, 14, 14] 2,304
Conv2d-547 [-1, 16, 14, 14] 2,304
Conv2d-548 [-1, 16, 14, 14] 2,304
Conv2d-549 [-1, 16, 14, 14] 2,304
Conv2d-550 [-1, 16, 14, 14] 2,304
Conv2d-551 [-1, 16, 14, 14] 2,304
Conv2d-552 [-1, 16, 14, 14] 2,304
Conv2d-553 [-1, 16, 14, 14] 2,304
Conv2d-554 [-1, 16, 14, 14] 2,304
Conv2d-555 [-1, 16, 14, 14] 2,304
Conv2d-556 [-1, 16, 14, 14] 2,304
Conv2d-557 [-1, 16, 14, 14] 2,304
Conv2d-558 [-1, 16, 14, 14] 2,304
BatchNorm2d-559 [-1, 512, 14, 14] 1,024
ReLU-560 [-1, 512, 14, 14] 0
GroupedConvBlock-561 [-1, 512, 14, 14] 0
Conv2d-562 [-1, 1024, 14, 14] 524,288
BatchNorm2d-563 [-1, 1024, 14, 14] 2,048
ReLU-564 [-1, 1024, 14, 14] 0
Block-565 [-1, 1024, 14, 14] 0
Stack-566 [-1, 1024, 14, 14] 0
Conv2d-567 [-1, 2048, 7, 7] 2,097,152
Conv2d-568 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-569 [-1, 1024, 14, 14] 2,048
ReLU-570 [-1, 1024, 14, 14] 0
Conv2d-571 [-1, 32, 7, 7] 9,216
Conv2d-572 [-1, 32, 7, 7] 9,216
Conv2d-573 [-1, 32, 7, 7] 9,216
Conv2d-574 [-1, 32, 7, 7] 9,216
Conv2d-575 [-1, 32, 7, 7] 9,216
Conv2d-576 [-1, 32, 7, 7] 9,216
Conv2d-577 [-1, 32, 7, 7] 9,216
Conv2d-578 [-1, 32, 7, 7] 9,216
Conv2d-579 [-1, 32, 7, 7] 9,216
Conv2d-580 [-1, 32, 7, 7] 9,216
Conv2d-581 [-1, 32, 7, 7] 9,216
Conv2d-582 [-1, 32, 7, 7] 9,216
Conv2d-583 [-1, 32, 7, 7] 9,216
Conv2d-584 [-1, 32, 7, 7] 9,216
Conv2d-585 [-1, 32, 7, 7] 9,216
Conv2d-586 [-1, 32, 7, 7] 9,216
Conv2d-587 [-1, 32, 7, 7] 9,216
Conv2d-588 [-1, 32, 7, 7] 9,216
Conv2d-589 [-1, 32, 7, 7] 9,216
Conv2d-590 [-1, 32, 7, 7] 9,216
Conv2d-591 [-1, 32, 7, 7] 9,216
Conv2d-592 [-1, 32, 7, 7] 9,216
Conv2d-593 [-1, 32, 7, 7] 9,216
Conv2d-594 [-1, 32, 7, 7] 9,216
Conv2d-595 [-1, 32, 7, 7] 9,216
Conv2d-596 [-1, 32, 7, 7] 9,216
Conv2d-597 [-1, 32, 7, 7] 9,216
Conv2d-598 [-1, 32, 7, 7] 9,216
Conv2d-599 [-1, 32, 7, 7] 9,216
Conv2d-600 [-1, 32, 7, 7] 9,216
Conv2d-601 [-1, 32, 7, 7] 9,216
Conv2d-602 [-1, 32, 7, 7] 9,216
BatchNorm2d-603 [-1, 1024, 7, 7] 2,048
ReLU-604 [-1, 1024, 7, 7] 0
GroupedConvBlock-605 [-1, 1024, 7, 7] 0
Conv2d-606 [-1, 2048, 7, 7] 2,097,152
BatchNorm2d-607 [-1, 2048, 7, 7] 4,096
ReLU-608 [-1, 2048, 7, 7] 0
Block-609 [-1, 2048, 7, 7] 0
Identity-610 [-1, 2048, 7, 7] 0
Conv2d-611 [-1, 1024, 7, 7] 2,097,152
BatchNorm2d-612 [-1, 1024, 7, 7] 2,048
ReLU-613 [-1, 1024, 7, 7] 0
Conv2d-614 [-1, 32, 7, 7] 9,216
Conv2d-615 [-1, 32, 7, 7] 9,216
Conv2d-616 [-1, 32, 7, 7] 9,216
Conv2d-617 [-1, 32, 7, 7] 9,216
Conv2d-618 [-1, 32, 7, 7] 9,216
Conv2d-619 [-1, 32, 7, 7] 9,216
Conv2d-620 [-1, 32, 7, 7] 9,216
Conv2d-621 [-1, 32, 7, 7] 9,216
Conv2d-622 [-1, 32, 7, 7] 9,216
Conv2d-623 [-1, 32, 7, 7] 9,216
Conv2d-624 [-1, 32, 7, 7] 9,216
Conv2d-625 [-1, 32, 7, 7] 9,216
Conv2d-626 [-1, 32, 7, 7] 9,216
Conv2d-627 [-1, 32, 7, 7] 9,216
Conv2d-628 [-1, 32, 7, 7] 9,216
Conv2d-629 [-1, 32, 7, 7] 9,216
Conv2d-630 [-1, 32, 7, 7] 9,216
Conv2d-631 [-1, 32, 7, 7] 9,216
Conv2d-632 [-1, 32, 7, 7] 9,216
Conv2d-633 [-1, 32, 7, 7] 9,216
Conv2d-634 [-1, 32, 7, 7] 9,216
Conv2d-635 [-1, 32, 7, 7] 9,216
Conv2d-636 [-1, 32, 7, 7] 9,216
Conv2d-637 [-1, 32, 7, 7] 9,216
Conv2d-638 [-1, 32, 7, 7] 9,216
Conv2d-639 [-1, 32, 7, 7] 9,216
Conv2d-640 [-1, 32, 7, 7] 9,216
Conv2d-641 [-1, 32, 7, 7] 9,216
Conv2d-642 [-1, 32, 7, 7] 9,216
Conv2d-643 [-1, 32, 7, 7] 9,216
Conv2d-644 [-1, 32, 7, 7] 9,216
Conv2d-645 [-1, 32, 7, 7] 9,216
BatchNorm2d-646 [-1, 1024, 7, 7] 2,048
ReLU-647 [-1, 1024, 7, 7] 0
GroupedConvBlock-648 [-1, 1024, 7, 7] 0
Conv2d-649 [-1, 2048, 7, 7] 2,097,152
BatchNorm2d-650 [-1, 2048, 7, 7] 4,096
ReLU-651 [-1, 2048, 7, 7] 0
Block-652 [-1, 2048, 7, 7] 0
Identity-653 [-1, 2048, 7, 7] 0
Conv2d-654 [-1, 1024, 7, 7] 2,097,152
BatchNorm2d-655 [-1, 1024, 7, 7] 2,048
ReLU-656 [-1, 1024, 7, 7] 0
Conv2d-657 [-1, 32, 7, 7] 9,216
Conv2d-658 [-1, 32, 7, 7] 9,216
Conv2d-659 [-1, 32, 7, 7] 9,216
Conv2d-660 [-1, 32, 7, 7] 9,216
Conv2d-661 [-1, 32, 7, 7] 9,216
Conv2d-662 [-1, 32, 7, 7] 9,216
Conv2d-663 [-1, 32, 7, 7] 9,216
Conv2d-664 [-1, 32, 7, 7] 9,216
Conv2d-665 [-1, 32, 7, 7] 9,216
Conv2d-666 [-1, 32, 7, 7] 9,216
Conv2d-667 [-1, 32, 7, 7] 9,216
Conv2d-668 [-1, 32, 7, 7] 9,216
Conv2d-669 [-1, 32, 7, 7] 9,216
Conv2d-670 [-1, 32, 7, 7] 9,216
Conv2d-671 [-1, 32, 7, 7] 9,216
Conv2d-672 [-1, 32, 7, 7] 9,216
Conv2d-673 [-1, 32, 7, 7] 9,216
Conv2d-674 [-1, 32, 7, 7] 9,216
Conv2d-675 [-1, 32, 7, 7] 9,216
Conv2d-676 [-1, 32, 7, 7] 9,216
Conv2d-677 [-1, 32, 7, 7] 9,216
Conv2d-678 [-1, 32, 7, 7] 9,216
Conv2d-679 [-1, 32, 7, 7] 9,216
Conv2d-680 [-1, 32, 7, 7] 9,216
Conv2d-681 [-1, 32, 7, 7] 9,216
Conv2d-682 [-1, 32, 7, 7] 9,216
Conv2d-683 [-1, 32, 7, 7] 9,216
Conv2d-684 [-1, 32, 7, 7] 9,216
Conv2d-685 [-1, 32, 7, 7] 9,216
Conv2d-686 [-1, 32, 7, 7] 9,216
Conv2d-687 [-1, 32, 7, 7] 9,216
Conv2d-688 [-1, 32, 7, 7] 9,216
BatchNorm2d-689 [-1, 1024, 7, 7] 2,048
ReLU-690 [-1, 1024, 7, 7] 0
GroupedConvBlock-691 [-1, 1024, 7, 7] 0
Conv2d-692 [-1, 2048, 7, 7] 2,097,152
BatchNorm2d-693 [-1, 2048, 7, 7] 4,096
ReLU-694 [-1, 2048, 7, 7] 0
Block-695 [-1, 2048, 7, 7] 0
Stack-696 [-1, 2048, 7, 7] 0
AdaptiveAvgPool2d-697 [-1, 2048, 1, 1] 0
Flatten-698 [-1, 2048] 0
Linear-699 [-1, 2] 4,098
================================================================
Total params: 22,976,386
Trainable params: 22,976,386
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 413.47
Params size (MB): 87.65
Estimated Total Size (MB): 501.69
----------------------------------------------------------------

ResNeXt50(
(conv1): Sequential(
(conv): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3))
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU()
(max_pool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
)
(conv2): Stack(
(conv): Sequential(
(0): Block(
(short): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(conv1): Sequential(
(0): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(1): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(2): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
)
)
(conv3): Stack(
(conv): Sequential(
(0): Block(
(short): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
(conv1): Sequential(
(0): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(8, 8, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(norm): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(1): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(2): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(3): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
)
)
(conv4): Stack(
(conv): Sequential(
(0): Block(
(short): Conv2d(512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False)
(conv1): Sequential(
(0): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(norm): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(1): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(2): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(3): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(4): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(5): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
)
)
(conv5): Stack(
(conv): Sequential(
(0): Block(
(short): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False)
(conv1): Sequential(
(0): Conv2d(1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(norm): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(1): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(2048, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
(2): Block(
(short): Identity()
(conv1): Sequential(
(0): Conv2d(2048, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
(conv2): GroupedConvBlock(
(conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(norm): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
)
(conv3): Sequential(
(0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(relu): ReLU(inplace=True)
)
)
)
(post): Sequential(
(avg_pool): AdaptiveAvgPool2d(output_size=(1, 1))
(flatten): Flatten(start_dim=1, end_dim=-1)
(fc): Linear(in_features=2048, out_features=2, bias=True)
)
)

4 总结

总而言之，ResNeXt是在ResNet的网络架构上，使用类似于Inception的分治思想，即split-tranform-merge策略，将模块中的网络拆开分组，与Inception不同，每组的卷积核大小一致，这样其感受野一致，但由于每组的卷积核参数不同，提取的特征自然不同。然后将每组得到的特征进行concat操作后，再与原输入特征x或者经过卷积等处理（即进行非线性变换）的特征进行Add操作。这样做的好处是，在不增加参数复杂度的前提下提高准确率，同时还能提高超参数的数量。

另外，cardinality是基的意思，将数个通道特征进行分组，不同的特征组之间可以看作是由不同基组成的子空间，每个组的核虽然一样，但参数不同，在各自的子空间中学到的特征就多种多样，这点跟transformer中的Multi-head attention不谋而合（Multi-head attention allows the model to jointly attend to information from different representation subspaces.）而且分组进行特征提取，使得学到的特征冗余度降低，获取能起到正则化的作用。