第J6周：ResNeXt-50实战
目录
前言
1.检查GPU
2.查看数据
3.划分数据集
4.创建模型
5.编译及训练模型
6.结果可视化
7.总结
前言
- 🍨 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
1.检查GPU
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
from torchvision import transforms, datasetsimport os,PIL,pathlibdevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")devicedevice(type='cpu')
2.查看数据
import os,PIL,random,pathlibdata_dir = '具体路径'
data_dir = pathlib.Path(data_dir)data_paths = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[2] for path in data_paths]
classeNames3.划分数据集
total_datadir = 'data/45-data'train_transforms = transforms.Compose([transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间transforms.Normalize( # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])total_data = datasets.ImageFolder(total_datadir,transform=train_transforms)
total_datatrain_size = int(0.8 * len(total_data))
test_size = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
train_dataset, test_datasettrain_size,test_sizebatch_size = 32train_dl = torch.utils.data.DataLoader(train_dataset,batch_size=batch_size,shuffle=True,num_workers=1)
test_dl = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=True,num_workers=1)for X, y in test_dl:print("Shape of X [N, C, H, W]: ", X.shape)print("Shape of y: ", y.shape, y.dtype)breakShape of X [N, C, H, W]: torch.Size([32, 3, 224, 224]) Shape of y: torch.Size([32]) torch.int64
4.创建模型
import torch
import torch.nn as nn
import torch.nn.functional as Fclass GroupedConvolutionBlock(nn.Module):def __init__(self, in_channels, out_channels, strides, groups):super(GroupedConvolutionBlock, self).__init__()self.groups = groupsself.g_channels = out_channels // groupsself.conv_layers = nn.ModuleList([nn.Conv2d(self.g_channels, self.g_channels, kernel_size=3, stride=strides, padding=1, bias=False)for _ in range(groups)])self.bn = nn.BatchNorm2d(out_channels, eps=1.001e-5)self.relu = nn.ReLU()def forward(self, x):group_list = []# 分组进行卷积for c in range(self.groups):# 分组取出数据x_group = x[:, c * self.g_channels:(c + 1) * self.g_channels, :, :]# 分组进行卷积x_group = self.conv_layers[c](x_group)# 存入listgroup_list.append(x_group)# 合并list中的数据group_merge = torch.cat(group_list, dim=1)x = self.bn(group_merge)x = self.relu(x)return xclass Block(nn.Module):def __init__(self, in_channels, filters, strides=1, groups=32, conv_shortcut=True):super(Block, self).__init__()self.conv_shortcut = conv_shortcutif conv_shortcut:self.shortcut = nn.Sequential(nn.Conv2d(in_channels, filters * 2, kernel_size=1, stride=strides, bias=False),nn.BatchNorm2d(filters * 2, eps=1.001e-5))else:self.shortcut = nn.Identity()self.conv1 = nn.Conv2d(in_channels, filters, kernel_size=1, stride=1, bias=False)self.bn1 = nn.BatchNorm2d(filters, eps=1.001e-5)self.relu1 = nn.ReLU()self.grouped_conv = GroupedConvolutionBlock(filters, filters, strides, groups)self.conv2 = nn.Conv2d(filters, filters * 2, kernel_size=1, stride=1, bias=False)self.bn2 = nn.BatchNorm2d(filters * 2, eps=1.001e-5)self.relu2 = nn.ReLU()def forward(self, x):shortcut = self.shortcut(x)x = self.conv1(x)x = self.bn1(x)x = self.relu1(x)x = self.grouped_conv(x)x = self.conv2(x)x = self.bn2(x)x = x + shortcutx = self.relu2(x)return xclass Stack(nn.Module):def __init__(self, in_channels, filters, blocks, strides, groups=32):super(Stack, self).__init__()self.blocks = nn.ModuleList()self.blocks.append(Block(in_channels, filters, strides, groups, conv_shortcut=True))for _ in range(1, blocks):self.blocks.append(Block(filters * 2, filters, strides=1, groups=groups, conv_shortcut=False))def forward(self, x):for block in self.blocks:x = block(x)return xclass ResNext50(nn.Module):def __init__(self, input_shape, num_classes):super(ResNext50, self).__init__()self.conv1 = nn.Conv2d(input_shape[0], 64, kernel_size=7, stride=2, padding=3, bias=False)self.bn1 = nn.BatchNorm2d(64, eps=1.001e-5)self.relu1 = nn.ReLU()self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)self.stack1 = Stack(64, 128, 2, 1)self.stack2 = Stack(256, 256, 3, 2)self.stack3 = Stack(512, 512, 5, 2)self.stack4 = Stack(1024, 1024, 2, 2)self.avgpool = nn.AdaptiveAvgPool2d((1, 1))self.fc = nn.Linear(2048, num_classes)def forward(self, x):x = self.conv1(x)x = self.bn1(x)x = self.relu1(x)x = self.maxpool(x)x = self.stack1(x)x = self.stack2(x)x = self.stack3(x)x = self.stack4(x)x = self.avgpool(x)x = torch.flatten(x, 1)x = self.fc(x)return xfrom torchsummary import summarymodel=ResNext50(input_shape=(224,224,3),num_classes=1000)model = ResNext50(input_shape=(3, 224, 224), num_classes=1000)# 将模型移动到GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)# 打印模型摘要
summary(model, input_size=(3, 224, 224))----------------------------------------------------------------Layer (type) Output Shape Param # ================================================================Conv2d-1 [-1, 64, 112, 112] 9,408BatchNorm2d-2 [-1, 64, 112, 112] 128ReLU-3 [-1, 64, 112, 112] 0MaxPool2d-4 [-1, 64, 56, 56] 0Conv2d-5 [-1, 256, 56, 56] 16,384BatchNorm2d-6 [-1, 256, 56, 56] 512Conv2d-7 [-1, 128, 56, 56] 8,192BatchNorm2d-8 [-1, 128, 56, 56] 256ReLU-9 [-1, 128, 56, 56] 0Conv2d-10 [-1, 4, 56, 56] 144Conv2d-11 [-1, 4, 56, 56] 144Conv2d-12 [-1, 4, 56, 56] 144Conv2d-13 [-1, 4, 56, 56] 144Conv2d-14 [-1, 4, 56, 56] 144Conv2d-15 [-1, 4, 56, 56] 144Conv2d-16 [-1, 4, 56, 56] 144Conv2d-17 [-1, 4, 56, 56] 144Conv2d-18 [-1, 4, 56, 56] 144Conv2d-19 [-1, 4, 56, 56] 144Conv2d-20 [-1, 4, 56, 56] 144Conv2d-21 [-1, 4, 56, 56] 144Conv2d-22 [-1, 4, 56, 56] 144Conv2d-23 [-1, 4, 56, 56] 144Conv2d-24 [-1, 4, 56, 56] 144Conv2d-25 [-1, 4, 56, 56] 144Conv2d-26 [-1, 4, 56, 56] 144Conv2d-27 [-1, 4, 56, 56] 144Conv2d-28 [-1, 4, 56, 56] 144Conv2d-29 [-1, 4, 56, 56] 144Conv2d-30 [-1, 4, 56, 56] 144Conv2d-31 [-1, 4, 56, 56] 144Conv2d-32 [-1, 4, 56, 56] 144Conv2d-33 [-1, 4, 56, 56] 144Conv2d-34 [-1, 4, 56, 56] 144Conv2d-35 [-1, 4, 56, 56] 144Conv2d-36 [-1, 4, 56, 56] 144Conv2d-37 [-1, 4, 56, 56] 144Conv2d-38 [-1, 4, 56, 56] 144Conv2d-39 [-1, 4, 56, 56] 144Conv2d-40 [-1, 4, 56, 56] 144Conv2d-41 [-1, 4, 56, 56] 144BatchNorm2d-42 [-1, 128, 56, 56] 256ReLU-43 [-1, 128, 56, 56] 0 GroupedConvolutionBlock-44 [-1, 128, 56, 56] 0Conv2d-45 [-1, 256, 56, 56] 32,768BatchNorm2d-46 [-1, 256, 56, 56] 512ReLU-47 [-1, 256, 56, 56] 0Block-48 [-1, 256, 56, 56] 0Identity-49 [-1, 256, 56, 56] 0Conv2d-50 [-1, 128, 56, 56] 32,768BatchNorm2d-51 [-1, 128, 56, 56] 256ReLU-52 [-1, 128, 56, 56] 0Conv2d-53 [-1, 4, 56, 56] 144Conv2d-54 [-1, 4, 56, 56] 144Conv2d-55 [-1, 4, 56, 56] 144Conv2d-56 [-1, 4, 56, 56] 144Conv2d-57 [-1, 4, 56, 56] 144Conv2d-58 [-1, 4, 56, 56] 144Conv2d-59 [-1, 4, 56, 56] 144Conv2d-60 [-1, 4, 56, 56] 144Conv2d-61 [-1, 4, 56, 56] 144Conv2d-62 [-1, 4, 56, 56] 144Conv2d-63 [-1, 4, 56, 56] 144Conv2d-64 [-1, 4, 56, 56] 144Conv2d-65 [-1, 4, 56, 56] 144Conv2d-66 [-1, 4, 56, 56] 144Conv2d-67 [-1, 4, 56, 56] 144Conv2d-68 [-1, 4, 56, 56] 144Conv2d-69 [-1, 4, 56, 56] 144Conv2d-70 [-1, 4, 56, 56] 144Conv2d-71 [-1, 4, 56, 56] 144Conv2d-72 [-1, 4, 56, 56] 144Conv2d-73 [-1, 4, 56, 56] 144Conv2d-74 [-1, 4, 56, 56] 144Conv2d-75 [-1, 4, 56, 56] 144Conv2d-76 [-1, 4, 56, 56] 144Conv2d-77 [-1, 4, 56, 56] 144Conv2d-78 [-1, 4, 56, 56] 144Conv2d-79 [-1, 4, 56, 56] 144Conv2d-80 [-1, 4, 56, 56] 144Conv2d-81 [-1, 4, 56, 56] 144Conv2d-82 [-1, 4, 56, 56] 144Conv2d-83 [-1, 4, 56, 56] 144Conv2d-84 [-1, 4, 56, 56] 144BatchNorm2d-85 [-1, 128, 56, 56] 256ReLU-86 [-1, 128, 56, 56] 0 GroupedConvolutionBlock-87 [-1, 128, 56, 56] 0Conv2d-88 [-1, 256, 56, 56] 32,768BatchNorm2d-89 [-1, 256, 56, 56] 512ReLU-90 [-1, 256, 56, 56] 0Block-91 [-1, 256, 56, 56] 0Stack-92 [-1, 256, 56, 56] 0Conv2d-93 [-1, 512, 28, 28] 131,072BatchNorm2d-94 [-1, 512, 28, 28] 1,024Conv2d-95 [-1, 256, 56, 56] 65,536BatchNorm2d-96 [-1, 256, 56, 56] 512ReLU-97 [-1, 256, 56, 56] 0Conv2d-98 [-1, 8, 28, 28] 576Conv2d-99 [-1, 8, 28, 28] 576Conv2d-100 [-1, 8, 28, 28] 576Conv2d-101 [-1, 8, 28, 28] 576Conv2d-102 [-1, 8, 28, 28] 576Conv2d-103 [-1, 8, 28, 28] 576Conv2d-104 [-1, 8, 28, 28] 576Conv2d-105 [-1, 8, 28, 28] 576Conv2d-106 [-1, 8, 28, 28] 576Conv2d-107 [-1, 8, 28, 28] 576Conv2d-108 [-1, 8, 28, 28] 576Conv2d-109 [-1, 8, 28, 28] 576Conv2d-110 [-1, 8, 28, 28] 576Conv2d-111 [-1, 8, 28, 28] 576Conv2d-112 [-1, 8, 28, 28] 576Conv2d-113 [-1, 8, 28, 28] 576Conv2d-114 [-1, 8, 28, 28] 576Conv2d-115 [-1, 8, 28, 28] 576Conv2d-116 [-1, 8, 28, 28] 576Conv2d-117 [-1, 8, 28, 28] 576Conv2d-118 [-1, 8, 28, 28] 576Conv2d-119 [-1, 8, 28, 28] 576Conv2d-120 [-1, 8, 28, 28] 576Conv2d-121 [-1, 8, 28, 28] 576Conv2d-122 [-1, 8, 28, 28] 576Conv2d-123 [-1, 8, 28, 28] 576Conv2d-124 [-1, 8, 28, 28] 576Conv2d-125 [-1, 8, 28, 28] 576Conv2d-126 [-1, 8, 28, 28] 576Conv2d-127 [-1, 8, 28, 28] 576Conv2d-128 [-1, 8, 28, 28] 576Conv2d-129 [-1, 8, 28, 28] 576BatchNorm2d-130 [-1, 256, 28, 28] 512ReLU-131 [-1, 256, 28, 28] 0 GroupedConvolutionBlock-132 [-1, 256, 28, 28] 0Conv2d-133 [-1, 512, 28, 28] 131,072BatchNorm2d-134 [-1, 512, 28, 28] 1,024ReLU-135 [-1, 512, 28, 28] 0Block-136 [-1, 512, 28, 28] 0Identity-137 [-1, 512, 28, 28] 0Conv2d-138 [-1, 256, 28, 28] 131,072BatchNorm2d-139 [-1, 256, 28, 28] 512ReLU-140 [-1, 256, 28, 28] 0Conv2d-141 [-1, 8, 28, 28] 576Conv2d-142 [-1, 8, 28, 28] 576Conv2d-143 [-1, 8, 28, 28] 576Conv2d-144 [-1, 8, 28, 28] 576Conv2d-145 [-1, 8, 28, 28] 576Conv2d-146 [-1, 8, 28, 28] 576Conv2d-147 [-1, 8, 28, 28] 576Conv2d-148 [-1, 8, 28, 28] 576Conv2d-149 [-1, 8, 28, 28] 576Conv2d-150 [-1, 8, 28, 28] 576Conv2d-151 [-1, 8, 28, 28] 576Conv2d-152 [-1, 8, 28, 28] 576Conv2d-153 [-1, 8, 28, 28] 576Conv2d-154 [-1, 8, 28, 28] 576Conv2d-155 [-1, 8, 28, 28] 576Conv2d-156 [-1, 8, 28, 28] 576Conv2d-157 [-1, 8, 28, 28] 576Conv2d-158 [-1, 8, 28, 28] 576Conv2d-159 [-1, 8, 28, 28] 576Conv2d-160 [-1, 8, 28, 28] 576Conv2d-161 [-1, 8, 28, 28] 576Conv2d-162 [-1, 8, 28, 28] 576Conv2d-163 [-1, 8, 28, 28] 576Conv2d-164 [-1, 8, 28, 28] 576Conv2d-165 [-1, 8, 28, 28] 576Conv2d-166 [-1, 8, 28, 28] 576Conv2d-167 [-1, 8, 28, 28] 576Conv2d-168 [-1, 8, 28, 28] 576Conv2d-169 [-1, 8, 28, 28] 576Conv2d-170 [-1, 8, 28, 28] 576Conv2d-171 [-1, 8, 28, 28] 576Conv2d-172 [-1, 8, 28, 28] 576BatchNorm2d-173 [-1, 256, 28, 28] 512ReLU-174 [-1, 256, 28, 28] 0 GroupedConvolutionBlock-175 [-1, 256, 28, 28] 0Conv2d-176 [-1, 512, 28, 28] 131,072BatchNorm2d-177 [-1, 512, 28, 28] 1,024ReLU-178 [-1, 512, 28, 28] 0Block-179 [-1, 512, 28, 28] 0Identity-180 [-1, 512, 28, 28] 0Conv2d-181 [-1, 256, 28, 28] 131,072BatchNorm2d-182 [-1, 256, 28, 28] 512ReLU-183 [-1, 256, 28, 28] 0Conv2d-184 [-1, 8, 28, 28] 576Conv2d-185 [-1, 8, 28, 28] 576Conv2d-186 [-1, 8, 28, 28] 576Conv2d-187 [-1, 8, 28, 28] 576Conv2d-188 [-1, 8, 28, 28] 576Conv2d-189 [-1, 8, 28, 28] 576Conv2d-190 [-1, 8, 28, 28] 576Conv2d-191 [-1, 8, 28, 28] 576Conv2d-192 [-1, 8, 28, 28] 576Conv2d-193 [-1, 8, 28, 28] 576Conv2d-194 [-1, 8, 28, 28] 576Conv2d-195 [-1, 8, 28, 28] 576Conv2d-196 [-1, 8, 28, 28] 576Conv2d-197 [-1, 8, 28, 28] 576Conv2d-198 [-1, 8, 28, 28] 576Conv2d-199 [-1, 8, 28, 28] 576Conv2d-200 [-1, 8, 28, 28] 576Conv2d-201 [-1, 8, 28, 28] 576Conv2d-202 [-1, 8, 28, 28] 576Conv2d-203 [-1, 8, 28, 28] 576Conv2d-204 [-1, 8, 28, 28] 576Conv2d-205 [-1, 8, 28, 28] 576Conv2d-206 [-1, 8, 28, 28] 576Conv2d-207 [-1, 8, 28, 28] 576Conv2d-208 [-1, 8, 28, 28] 576Conv2d-209 [-1, 8, 28, 28] 576Conv2d-210 [-1, 8, 28, 28] 576Conv2d-211 [-1, 8, 28, 28] 576Conv2d-212 [-1, 8, 28, 28] 576Conv2d-213 [-1, 8, 28, 28] 576Conv2d-214 [-1, 8, 28, 28] 576Conv2d-215 [-1, 8, 28, 28] 576BatchNorm2d-216 [-1, 256, 28, 28] 512ReLU-217 [-1, 256, 28, 28] 0 GroupedConvolutionBlock-218 [-1, 256, 28, 28] 0Conv2d-219 [-1, 512, 28, 28] 131,072BatchNorm2d-220 [-1, 512, 28, 28] 1,024ReLU-221 [-1, 512, 28, 28] 0Block-222 [-1, 512, 28, 28] 0Stack-223 [-1, 512, 28, 28] 0Conv2d-224 [-1, 1024, 14, 14] 524,288BatchNorm2d-225 [-1, 1024, 14, 14] 2,048Conv2d-226 [-1, 512, 28, 28] 262,144BatchNorm2d-227 [-1, 512, 28, 28] 1,024ReLU-228 [-1, 512, 28, 28] 0Conv2d-229 [-1, 16, 14, 14] 2,304Conv2d-230 [-1, 16, 14, 14] 2,304Conv2d-231 [-1, 16, 14, 14] 2,304Conv2d-232 [-1, 16, 14, 14] 2,304Conv2d-233 [-1, 16, 14, 14] 2,304Conv2d-234 [-1, 16, 14, 14] 2,304Conv2d-235 [-1, 16, 14, 14] 2,304Conv2d-236 [-1, 16, 14, 14] 2,304Conv2d-237 [-1, 16, 14, 14] 2,304Conv2d-238 [-1, 16, 14, 14] 2,304Conv2d-239 [-1, 16, 14, 14] 2,304Conv2d-240 [-1, 16, 14, 14] 2,304Conv2d-241 [-1, 16, 14, 14] 2,304Conv2d-242 [-1, 16, 14, 14] 2,304Conv2d-243 [-1, 16, 14, 14] 2,304Conv2d-244 [-1, 16, 14, 14] 2,304Conv2d-245 [-1, 16, 14, 14] 2,304Conv2d-246 [-1, 16, 14, 14] 2,304Conv2d-247 [-1, 16, 14, 14] 2,304Conv2d-248 [-1, 16, 14, 14] 2,304Conv2d-249 [-1, 16, 14, 14] 2,304Conv2d-250 [-1, 16, 14, 14] 2,304Conv2d-251 [-1, 16, 14, 14] 2,304Conv2d-252 [-1, 16, 14, 14] 2,304Conv2d-253 [-1, 16, 14, 14] 2,304Conv2d-254 [-1, 16, 14, 14] 2,304Conv2d-255 [-1, 16, 14, 14] 2,304Conv2d-256 [-1, 16, 14, 14] 2,304Conv2d-257 [-1, 16, 14, 14] 2,304Conv2d-258 [-1, 16, 14, 14] 2,304Conv2d-259 [-1, 16, 14, 14] 2,304Conv2d-260 [-1, 16, 14, 14] 2,304BatchNorm2d-261 [-1, 512, 14, 14] 1,024ReLU-262 [-1, 512, 14, 14] 0 GroupedConvolutionBlock-263 [-1, 512, 14, 14] 0Conv2d-264 [-1, 1024, 14, 14] 524,288BatchNorm2d-265 [-1, 1024, 14, 14] 2,048ReLU-266 [-1, 1024, 14, 14] 0Block-267 [-1, 1024, 14, 14] 0Identity-268 [-1, 1024, 14, 14] 0Conv2d-269 [-1, 512, 14, 14] 524,288BatchNorm2d-270 [-1, 512, 14, 14] 1,024ReLU-271 [-1, 512, 14, 14] 0Conv2d-272 [-1, 16, 14, 14] 2,304Conv2d-273 [-1, 16, 14, 14] 2,304Conv2d-274 [-1, 16, 14, 14] 2,304Conv2d-275 [-1, 16, 14, 14] 2,304Conv2d-276 [-1, 16, 14, 14] 2,304Conv2d-277 [-1, 16, 14, 14] 2,304Conv2d-278 [-1, 16, 14, 14] 2,304Conv2d-279 [-1, 16, 14, 14] 2,304Conv2d-280 [-1, 16, 14, 14] 2,304Conv2d-281 [-1, 16, 14, 14] 2,304Conv2d-282 [-1, 16, 14, 14] 2,304Conv2d-283 [-1, 16, 14, 14] 2,304Conv2d-284 [-1, 16, 14, 14] 2,304Conv2d-285 [-1, 16, 14, 14] 2,304Conv2d-286 [-1, 16, 14, 14] 2,304Conv2d-287 [-1, 16, 14, 14] 2,304Conv2d-288 [-1, 16, 14, 14] 2,304Conv2d-289 [-1, 16, 14, 14] 2,304Conv2d-290 [-1, 16, 14, 14] 2,304Conv2d-291 [-1, 16, 14, 14] 2,304Conv2d-292 [-1, 16, 14, 14] 2,304Conv2d-293 [-1, 16, 14, 14] 2,304Conv2d-294 [-1, 16, 14, 14] 2,304Conv2d-295 [-1, 16, 14, 14] 2,304Conv2d-296 [-1, 16, 14, 14] 2,304Conv2d-297 [-1, 16, 14, 14] 2,304Conv2d-298 [-1, 16, 14, 14] 2,304Conv2d-299 [-1, 16, 14, 14] 2,304Conv2d-300 [-1, 16, 14, 14] 2,304Conv2d-301 [-1, 16, 14, 14] 2,304Conv2d-302 [-1, 16, 14, 14] 2,304Conv2d-303 [-1, 16, 14, 14] 2,304BatchNorm2d-304 [-1, 512, 14, 14] 1,024ReLU-305 [-1, 512, 14, 14] 0 GroupedConvolutionBlock-306 [-1, 512, 14, 14] 0Conv2d-307 [-1, 1024, 14, 14] 524,288BatchNorm2d-308 [-1, 1024, 14, 14] 2,048ReLU-309 [-1, 1024, 14, 14] 0Block-310 [-1, 1024, 14, 14] 0Identity-311 [-1, 1024, 14, 14] 0Conv2d-312 [-1, 512, 14, 14] 524,288BatchNorm2d-313 [-1, 512, 14, 14] 1,024ReLU-314 [-1, 512, 14, 14] 0Conv2d-315 [-1, 16, 14, 14] 2,304Conv2d-316 [-1, 16, 14, 14] 2,304Conv2d-317 [-1, 16, 14, 14] 2,304Conv2d-318 [-1, 16, 14, 14] 2,304Conv2d-319 [-1, 16, 14, 14] 2,304Conv2d-320 [-1, 16, 14, 14] 2,304Conv2d-321 [-1, 16, 14, 14] 2,304Conv2d-322 [-1, 16, 14, 14] 2,304Conv2d-323 [-1, 16, 14, 14] 2,304Conv2d-324 [-1, 16, 14, 14] 2,304Conv2d-325 [-1, 16, 14, 14] 2,304Conv2d-326 [-1, 16, 14, 14] 2,304Conv2d-327 [-1, 16, 14, 14] 2,304Conv2d-328 [-1, 16, 14, 14] 2,304Conv2d-329 [-1, 16, 14, 14] 2,304Conv2d-330 [-1, 16, 14, 14] 2,304Conv2d-331 [-1, 16, 14, 14] 2,304Conv2d-332 [-1, 16, 14, 14] 2,304Conv2d-333 [-1, 16, 14, 14] 2,304Conv2d-334 [-1, 16, 14, 14] 2,304Conv2d-335 [-1, 16, 14, 14] 2,304Conv2d-336 [-1, 16, 14, 14] 2,304Conv2d-337 [-1, 16, 14, 14] 2,304Conv2d-338 [-1, 16, 14, 14] 2,304Conv2d-339 [-1, 16, 14, 14] 2,304Conv2d-340 [-1, 16, 14, 14] 2,304Conv2d-341 [-1, 16, 14, 14] 2,304Conv2d-342 [-1, 16, 14, 14] 2,304Conv2d-343 [-1, 16, 14, 14] 2,304Conv2d-344 [-1, 16, 14, 14] 2,304Conv2d-345 [-1, 16, 14, 14] 2,304Conv2d-346 [-1, 16, 14, 14] 2,304BatchNorm2d-347 [-1, 512, 14, 14] 1,024ReLU-348 [-1, 512, 14, 14] 0 GroupedConvolutionBlock-349 [-1, 512, 14, 14] 0Conv2d-350 [-1, 1024, 14, 14] 524,288BatchNorm2d-351 [-1, 1024, 14, 14] 2,048ReLU-352 [-1, 1024, 14, 14] 0Block-353 [-1, 1024, 14, 14] 0Identity-354 [-1, 1024, 14, 14] 0Conv2d-355 [-1, 512, 14, 14] 524,288BatchNorm2d-356 [-1, 512, 14, 14] 1,024ReLU-357 [-1, 512, 14, 14] 0Conv2d-358 [-1, 16, 14, 14] 2,304Conv2d-359 [-1, 16, 14, 14] 2,304Conv2d-360 [-1, 16, 14, 14] 2,304Conv2d-361 [-1, 16, 14, 14] 2,304Conv2d-362 [-1, 16, 14, 14] 2,304Conv2d-363 [-1, 16, 14, 14] 2,304Conv2d-364 [-1, 16, 14, 14] 2,304Conv2d-365 [-1, 16, 14, 14] 2,304Conv2d-366 [-1, 16, 14, 14] 2,304Conv2d-367 [-1, 16, 14, 14] 2,304Conv2d-368 [-1, 16, 14, 14] 2,304Conv2d-369 [-1, 16, 14, 14] 2,304Conv2d-370 [-1, 16, 14, 14] 2,304Conv2d-371 [-1, 16, 14, 14] 2,304Conv2d-372 [-1, 16, 14, 14] 2,304Conv2d-373 [-1, 16, 14, 14] 2,304Conv2d-374 [-1, 16, 14, 14] 2,304Conv2d-375 [-1, 16, 14, 14] 2,304Conv2d-376 [-1, 16, 14, 14] 2,304Conv2d-377 [-1, 16, 14, 14] 2,304Conv2d-378 [-1, 16, 14, 14] 2,304Conv2d-379 [-1, 16, 14, 14] 2,304Conv2d-380 [-1, 16, 14, 14] 2,304Conv2d-381 [-1, 16, 14, 14] 2,304Conv2d-382 [-1, 16, 14, 14] 2,304Conv2d-383 [-1, 16, 14, 14] 2,304Conv2d-384 [-1, 16, 14, 14] 2,304Conv2d-385 [-1, 16, 14, 14] 2,304Conv2d-386 [-1, 16, 14, 14] 2,304Conv2d-387 [-1, 16, 14, 14] 2,304Conv2d-388 [-1, 16, 14, 14] 2,304Conv2d-389 [-1, 16, 14, 14] 2,304BatchNorm2d-390 [-1, 512, 14, 14] 1,024ReLU-391 [-1, 512, 14, 14] 0 GroupedConvolutionBlock-392 [-1, 512, 14, 14] 0Conv2d-393 [-1, 1024, 14, 14] 524,288BatchNorm2d-394 [-1, 1024, 14, 14] 2,048ReLU-395 [-1, 1024, 14, 14] 0Block-396 [-1, 1024, 14, 14] 0Identity-397 [-1, 1024, 14, 14] 0Conv2d-398 [-1, 512, 14, 14] 524,288BatchNorm2d-399 [-1, 512, 14, 14] 1,024ReLU-400 [-1, 512, 14, 14] 0Conv2d-401 [-1, 16, 14, 14] 2,304Conv2d-402 [-1, 16, 14, 14] 2,304Conv2d-403 [-1, 16, 14, 14] 2,304Conv2d-404 [-1, 16, 14, 14] 2,304Conv2d-405 [-1, 16, 14, 14] 2,304Conv2d-406 [-1, 16, 14, 14] 2,304Conv2d-407 [-1, 16, 14, 14] 2,304Conv2d-408 [-1, 16, 14, 14] 2,304Conv2d-409 [-1, 16, 14, 14] 2,304Conv2d-410 [-1, 16, 14, 14] 2,304Conv2d-411 [-1, 16, 14, 14] 2,304Conv2d-412 [-1, 16, 14, 14] 2,304Conv2d-413 [-1, 16, 14, 14] 2,304Conv2d-414 [-1, 16, 14, 14] 2,304Conv2d-415 [-1, 16, 14, 14] 2,304Conv2d-416 [-1, 16, 14, 14] 2,304Conv2d-417 [-1, 16, 14, 14] 2,304Conv2d-418 [-1, 16, 14, 14] 2,304Conv2d-419 [-1, 16, 14, 14] 2,304Conv2d-420 [-1, 16, 14, 14] 2,304Conv2d-421 [-1, 16, 14, 14] 2,304Conv2d-422 [-1, 16, 14, 14] 2,304Conv2d-423 [-1, 16, 14, 14] 2,304Conv2d-424 [-1, 16, 14, 14] 2,304Conv2d-425 [-1, 16, 14, 14] 2,304Conv2d-426 [-1, 16, 14, 14] 2,304Conv2d-427 [-1, 16, 14, 14] 2,304Conv2d-428 [-1, 16, 14, 14] 2,304Conv2d-429 [-1, 16, 14, 14] 2,304Conv2d-430 [-1, 16, 14, 14] 2,304Conv2d-431 [-1, 16, 14, 14] 2,304Conv2d-432 [-1, 16, 14, 14] 2,304BatchNorm2d-433 [-1, 512, 14, 14] 1,024ReLU-434 [-1, 512, 14, 14] 0 GroupedConvolutionBlock-435 [-1, 512, 14, 14] 0Conv2d-436 [-1, 1024, 14, 14] 524,288BatchNorm2d-437 [-1, 1024, 14, 14] 2,048ReLU-438 [-1, 1024, 14, 14] 0Block-439 [-1, 1024, 14, 14] 0Stack-440 [-1, 1024, 14, 14] 0Conv2d-441 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-442 [-1, 2048, 7, 7] 4,096Conv2d-443 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-444 [-1, 1024, 14, 14] 2,048ReLU-445 [-1, 1024, 14, 14] 0Conv2d-446 [-1, 32, 7, 7] 9,216Conv2d-447 [-1, 32, 7, 7] 9,216Conv2d-448 [-1, 32, 7, 7] 9,216Conv2d-449 [-1, 32, 7, 7] 9,216Conv2d-450 [-1, 32, 7, 7] 9,216Conv2d-451 [-1, 32, 7, 7] 9,216Conv2d-452 [-1, 32, 7, 7] 9,216Conv2d-453 [-1, 32, 7, 7] 9,216Conv2d-454 [-1, 32, 7, 7] 9,216Conv2d-455 [-1, 32, 7, 7] 9,216Conv2d-456 [-1, 32, 7, 7] 9,216Conv2d-457 [-1, 32, 7, 7] 9,216Conv2d-458 [-1, 32, 7, 7] 9,216Conv2d-459 [-1, 32, 7, 7] 9,216Conv2d-460 [-1, 32, 7, 7] 9,216Conv2d-461 [-1, 32, 7, 7] 9,216Conv2d-462 [-1, 32, 7, 7] 9,216Conv2d-463 [-1, 32, 7, 7] 9,216Conv2d-464 [-1, 32, 7, 7] 9,216Conv2d-465 [-1, 32, 7, 7] 9,216Conv2d-466 [-1, 32, 7, 7] 9,216Conv2d-467 [-1, 32, 7, 7] 9,216Conv2d-468 [-1, 32, 7, 7] 9,216Conv2d-469 [-1, 32, 7, 7] 9,216Conv2d-470 [-1, 32, 7, 7] 9,216Conv2d-471 [-1, 32, 7, 7] 9,216Conv2d-472 [-1, 32, 7, 7] 9,216Conv2d-473 [-1, 32, 7, 7] 9,216Conv2d-474 [-1, 32, 7, 7] 9,216Conv2d-475 [-1, 32, 7, 7] 9,216Conv2d-476 [-1, 32, 7, 7] 9,216Conv2d-477 [-1, 32, 7, 7] 9,216BatchNorm2d-478 [-1, 1024, 7, 7] 2,048ReLU-479 [-1, 1024, 7, 7] 0 GroupedConvolutionBlock-480 [-1, 1024, 7, 7] 0Conv2d-481 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-482 [-1, 2048, 7, 7] 4,096ReLU-483 [-1, 2048, 7, 7] 0Block-484 [-1, 2048, 7, 7] 0Identity-485 [-1, 2048, 7, 7] 0Conv2d-486 [-1, 1024, 7, 7] 2,097,152BatchNorm2d-487 [-1, 1024, 7, 7] 2,048ReLU-488 [-1, 1024, 7, 7] 0Conv2d-489 [-1, 32, 7, 7] 9,216Conv2d-490 [-1, 32, 7, 7] 9,216Conv2d-491 [-1, 32, 7, 7] 9,216Conv2d-492 [-1, 32, 7, 7] 9,216Conv2d-493 [-1, 32, 7, 7] 9,216Conv2d-494 [-1, 32, 7, 7] 9,216Conv2d-495 [-1, 32, 7, 7] 9,216Conv2d-496 [-1, 32, 7, 7] 9,216Conv2d-497 [-1, 32, 7, 7] 9,216Conv2d-498 [-1, 32, 7, 7] 9,216Conv2d-499 [-1, 32, 7, 7] 9,216Conv2d-500 [-1, 32, 7, 7] 9,216Conv2d-501 [-1, 32, 7, 7] 9,216Conv2d-502 [-1, 32, 7, 7] 9,216Conv2d-503 [-1, 32, 7, 7] 9,216Conv2d-504 [-1, 32, 7, 7] 9,216Conv2d-505 [-1, 32, 7, 7] 9,216Conv2d-506 [-1, 32, 7, 7] 9,216Conv2d-507 [-1, 32, 7, 7] 9,216Conv2d-508 [-1, 32, 7, 7] 9,216Conv2d-509 [-1, 32, 7, 7] 9,216Conv2d-510 [-1, 32, 7, 7] 9,216Conv2d-511 [-1, 32, 7, 7] 9,216Conv2d-512 [-1, 32, 7, 7] 9,216Conv2d-513 [-1, 32, 7, 7] 9,216Conv2d-514 [-1, 32, 7, 7] 9,216Conv2d-515 [-1, 32, 7, 7] 9,216Conv2d-516 [-1, 32, 7, 7] 9,216Conv2d-517 [-1, 32, 7, 7] 9,216Conv2d-518 [-1, 32, 7, 7] 9,216Conv2d-519 [-1, 32, 7, 7] 9,216Conv2d-520 [-1, 32, 7, 7] 9,216BatchNorm2d-521 [-1, 1024, 7, 7] 2,048ReLU-522 [-1, 1024, 7, 7] 0 GroupedConvolutionBlock-523 [-1, 1024, 7, 7] 0Conv2d-524 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-525 [-1, 2048, 7, 7] 4,096ReLU-526 [-1, 2048, 7, 7] 0Block-527 [-1, 2048, 7, 7] 0Stack-528 [-1, 2048, 7, 7] 0 AdaptiveAvgPool2d-529 [-1, 2048, 1, 1] 0Linear-530 [-1, 1000] 2,049,000 ================================================================ Total params: 19,051,304 Trainable params: 19,051,304 Non-trainable params: 0 ---------------------------------------------------------------- Input size (MB): 0.57 Forward/backward pass size (MB): 327.33 Params size (MB): 72.67 Estimated Total Size (MB): 400.58 ----------------------------------------------------------------
5.编译及训练模型
loss_fn = nn.CrossEntropyLoss() # 创建损失函数
learn_rate = 1e-4 # 学习率
opt = torch.optim.SGD(model.parameters(),lr=learn_rate)# 训练循环
def train(dataloader, model, loss_fn, optimizer):size = len(dataloader.dataset) # 训练集的大小，一共60000张图片num_batches = len(dataloader) # 批次数目，1875（60000/32）train_loss, train_acc = 0, 0 # 初始化训练损失和正确率for X, y in dataloader: # 获取图片及其标签X, y = X.to(device), y.to(device)# 计算预测误差pred = model(X) # 网络输出loss = loss_fn(pred, y) # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失# 反向传播optimizer.zero_grad() # grad属性归零loss.backward() # 反向传播optimizer.step() # 每一步自动更新# 记录acc与losstrain_acc += (pred.argmax(1) == y).type(torch.float).sum().item()train_loss += loss.item()train_acc /= sizetrain_loss /= num_batchesreturn train_acc, train_lossdef test (dataloader, model, loss_fn):size = len(dataloader.dataset) # 测试集的大小，一共10000张图片num_batches = len(dataloader) # 批次数目，313（10000/32=312.5，向上取整）test_loss, test_acc = 0, 0# 当不进行训练时，停止梯度更新，节省计算内存消耗with torch.no_grad():for imgs, target in dataloader:imgs, target = imgs.to(device), target.to(device)# 计算losstarget_pred = model(imgs)loss = loss_fn(target_pred, target)test_loss += loss.item()test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()test_acc /= sizetest_loss /= num_batchesreturn test_acc, test_lossepochs = 20
train_loss = []
train_acc = []
test_loss = []
test_acc = []for epoch in range(epochs):model.train()epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, opt)model.eval()epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)train_acc.append(epoch_train_acc)train_loss.append(epoch_train_loss)test_acc.append(epoch_test_acc)test_loss.append(epoch_test_loss)template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%，Test_loss:{:.3f}')print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss, epoch_test_acc*100, epoch_test_loss))
print('Done')Epoch: 1, Train_acc:35.4%, Train_loss:5.256, Test_acc:53.4%，Test_loss:4.276 Epoch: 2, Train_acc:55.5%, Train_loss:2.817, Test_acc:52.0%，Test_loss:2.263。。。。。。
6.结果可视化
import matplotlib.pyplot as plt
#隐藏警告
import warnings
warnings.filterwarnings("ignore") #忽略警告信息
plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
plt.rcParams['figure.dpi'] = 100 #分辨率from datetime import datetime
current_time = datetime.now() # 获取当前时间epochs_range = range(epochs)plt.figure(figsize=(12, 3))
plt.subplot(1, 2, 1)plt.plot(epochs_range, train_acc, label='Training Accuracy')
plt.plot(epochs_range, test_acc, label='Test Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.xlabel(current_time) # 打卡请带上时间戳，否则代码截图无效plt.subplot(1, 2, 2)
plt.plot(epochs_range, train_loss, label='Training Loss')
plt.plot(epochs_range, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
7.总结
这次的实验，我选择将 SE（Squeeze-and-Excitation）模块集成到 ResNeXt-50 网络中，以探索其对模型表达能力和分类精度的提升效果。
ResNeXt-50 是一种融合了 ResNet 的残差连接思想与 Inception 式多分支结构的网络架构，其核心思想在于引入 Cardinality（基数），即增加“路径的数目”而不是仅堆叠更多的层或通道数，从而以较低计算成本获得更强的表现能力。在这种结构中，每一个 bottleneck 块本质上是多个分支的聚合输出，这为嵌入注意力机制提供了良好的基础。
我将 SE 模块嵌入到每一个 ResNeXt bottleneck 的输出之后，即在残差加和操作之前先进行通道注意力加权。这样做的原因是，SE 模块会根据当前特征的全局信息重新调整每个通道的权重，而 residual 加法操作是在全通道层面进行的。因此，先引入 SE 模块可以使注意力机制更直接地作用于主分支的特征流。
SE 模块本身的结构简单但有效：通过全局平均池化压缩空间维度，生成一个长度为 C 的向量（C 为通道数），然后依次通过两个全连接层和一个 Sigmoid 激活函数，生成通道权重系数。这个权重再与原特征图进行逐通道相乘，从而实现“强调重要、抑制冗余”的特征调整。
从实验结果来看，加入 SE 模块后的 ResNeXt-50 在多个标准数据集（如 CIFAR-100、ImageNet 子集）上表现出更加稳定的收敛速度和略微提升的最终精度。这种提升尤其体现在模型对细粒度类别的识别上，说明 SE 模块在强调重要通道特征方面确实发挥了作用。
整体来看，将 SE 模块应用于 ResNeXt-50 是一种结构上自然、实现上高效、效果上稳健的改进路径。它充分利用了 ResNeXt 的多分支表达能力和 SE 的通道动态建模机制，是一种值得借鉴的融合策略。