- 🍨 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
ResNeXt-50实战
# 定义transforms
import torchvision.transforms as transformstrain_transforms = transforms.Compose([transforms.Resize([224, 224]), # 统一图片尺寸transforms.RandomHorizontalFlip(p=0.5), # 随机水平翻转transforms.RandomRotation(degrees=15), # 小角度随机旋转transforms.RandomResizedCrop(size=224, scale=(0.8, 1.2)), # 随机缩放裁剪transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.1), # 光照变化transforms.ToTensor(), # 转换为 Tensor 格式transforms.Normalize( # 标准化mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
])
import torch
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
import warningswarnings.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
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
data_dir = r"C:\Users\11054\Desktop\kLearning\J1_learning\bird_photos"transform = transforms.Compose([transforms.Resize((224, 224)),transforms.ToTensor(),
])train_dataset = datasets.ImageFolder(data_dir, transform=transform)
train_size = int(0.8 * len(train_dataset))
val_size = len(train_dataset) - train_size
train_dataset, val_dataset = torch.utils.data.random_split(train_dataset, [train_size, val_size])train_loader = DataLoader(train_dataset, batch_size=8, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=8, shuffle=False)class_names = train_dataset.dataset.classes
class_names
['Bananaquit', 'Black Skimmer', 'Black Throated Bushtiti', 'Cockatoo']
train_transforms = transforms.Compose([transforms.Resize([224, 224]), # 统一图片尺寸transforms.RandomHorizontalFlip(p=0.5), # 随机水平翻转transforms.RandomRotation(degrees=15), # 小角度随机旋转transforms.RandomResizedCrop(size=224, scale=(0.8, 1.2)), # 随机缩放裁剪transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.1), # 光照变化transforms.ToTensor(), # 转换为 Tensor 格式transforms.Normalize( # 标准化mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
])test_transforms = transforms.Compose([transforms.Resize([224, 224]),transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
])
total_data = datasets.ImageFolder(data_dir, transform=train_transforms)
total_data
Dataset ImageFolderNumber of datapoints: 565Root location: C:\Users\11054\Desktop\kLearning\J1_learning\bird_photosStandardTransform
Transform: Compose(Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=True)RandomHorizontalFlip(p=0.5)RandomRotation(degrees=[-15.0, 15.0], interpolation=nearest, expand=False, fill=0)RandomResizedCrop(size=(224, 224), scale=(0.8, 1.2), ratio=(0.75, 1.3333), interpolation=bilinear, antialias=True)ColorJitter(brightness=(0.8, 1.2), contrast=(0.8, 1.2), saturation=(0.9, 1.1), hue=None)ToTensor()Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]))
# 设置随机种子
torch.manual_seed(42)# 划分数据集
train_size = int(len(total_data) * 0.8)
test_size = len(total_data) - train_sizetrain_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
train_dataset, test_dataset(<torch.utils.data.dataset.Subset at 0x184276162b0>,<torch.utils.data.dataset.Subset at 0x184276163a0>)
# 定义DataLoader用于数据集的加载batch_size = 8train_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
)
# 观察数据维度
for X, y in test_dl:print("Shape of X [N,C,H,W]: ", X.shape)print("Shape of y: ", y.shape, y.dtype)break
Shape of X [N,C,H,W]: torch.Size([8, 3, 224, 224])
Shape of y: torch.Size([8]) torch.int64
import torch
import torch.nn as nn# 定义分组卷积模块
class GroupedConvBlock(nn.Module):def __init__(self, in_channels, groups, g_channels, stride):super(GroupedConvBlock, self).__init__()self.groups = groupsself.group_conv = nn.ModuleList([nn.Conv2d(g_channels, g_channels, kernel_size=3, stride=stride, padding=1, bias=False)for _ in range(groups)])self.bn = nn.BatchNorm2d(in_channels)self.relu = nn.ReLU(inplace=True)def forward(self, x):# 分组数据split_x = torch.split(x, x.size(1) // self.groups, dim=1)group_out = [conv(g) for g, conv in zip(split_x, self.group_conv)]# 合并数据x = torch.cat(group_out, dim=1)x = self.bn(x)x = self.relu(x)return x# 定义残差模块
class ResNeXtBlock(nn.Module):def __init__(self, in_channels, filters, groups=32, stride=1, conv_shortcut=True):super(ResNeXtBlock, self).__init__()self.conv_shortcut = conv_shortcutself.groups = groupsself.g_channels = filters // groups# Shortcut分支if conv_shortcut:self.shortcut = nn.Sequential(nn.Conv2d(in_channels, filters * 2, kernel_size=1, stride=stride, bias=False),nn.BatchNorm2d(filters * 2),)else:self.shortcut = nn.Identity()# 主分支self.conv1 = nn.Sequential(nn.Conv2d(in_channels, filters, kernel_size=1, stride=1, bias=False),nn.BatchNorm2d(filters),nn.ReLU(inplace=True))self.grouped_conv = GroupedConvBlock(filters, groups, self.g_channels, stride)self.conv3 = nn.Sequential(nn.Conv2d(filters, filters * 2, kernel_size=1, stride=1, bias=False),nn.BatchNorm2d(filters * 2),)self.relu = nn.ReLU(inplace=True)def forward(self, x):shortcut = self.shortcut(x)x = self.conv1(x)x = self.grouped_conv(x)x = self.conv3(x)x += shortcutx = self.relu(x)return x# 定义 ResNeXt-50 模型
class ResNeXt50(nn.Module):def __init__(self, num_classes=1000):super(ResNeXt50, self).__init__()self.stem = nn.Sequential(nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),nn.BatchNorm2d(64),nn.ReLU(inplace=True),nn.MaxPool2d(kernel_size=3, stride=2, padding=1))# 堆叠ResNeXt模块self.layer1 = self.make_layer(64, 128, 3, stride=1)self.layer2 = self.make_layer(256, 256, 4, stride=2)self.layer3 = self.make_layer(512, 512, 6, stride=2)self.layer4 = self.make_layer(1024, 1024, 3, stride=2)# 全局平均池化和分类层self.global_avg_pool = nn.AdaptiveAvgPool2d(1)self.fc = nn.Linear(2048, num_classes)@staticmethoddef make_layer(in_channels, filters, blocks, stride):layers = [ResNeXtBlock(in_channels, filters, stride=stride)]for _ in range(1, blocks):layers.append(ResNeXtBlock(filters * 2, filters, stride=1))return nn.Sequential(*layers)def forward(self, x):x = self.stem(x)x = self.layer1(x)x = self.layer2(x)x = self.layer3(x)x = self.layer4(x)x = self.global_avg_pool(x)x = torch.flatten(x, 1)x = self.fc(x)return x
model = ResNeXt50(num_classes=4).to(device)
model
ResNeXt50((stem): Sequential((0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(2): ReLU(inplace=True)(3): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False))(layer1): Sequential((0): ResNeXtBlock((shortcut): Sequential((0): Conv2d(64, 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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)))(layer2): Sequential((0): ResNeXtBlock((shortcut): Sequential((0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(8, 8, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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)))(layer3): Sequential((0): ResNeXtBlock((shortcut): Sequential((0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(16, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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)))(layer4): Sequential((0): ResNeXtBlock((shortcut): Sequential((0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(32, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): Sequential((0): Conv2d(2048, 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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): ResNeXtBlock((shortcut): Sequential((0): Conv2d(2048, 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))(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))(grouped_conv): GroupedConvBlock((group_convs): ModuleList((0-31): 32 x Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(bn): 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)))(global_avg_pool): AdaptiveAvgPool2d(output_size=1)(fc): Linear(in_features=2048, out_features=4, bias=True)
)
# 查看模型详情
import torchsummary as summarysummary.summary(model, (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] 0GroupedConvBlock-44 [-1, 128, 56, 56] 0Conv2d-45 [-1, 256, 56, 56] 32,768BatchNorm2d-46 [-1, 256, 56, 56] 512ReLU-47 [-1, 256, 56, 56] 0ResNeXtBlock-48 [-1, 256, 56, 56] 0Conv2d-49 [-1, 256, 56, 56] 65,536BatchNorm2d-50 [-1, 256, 56, 56] 512Conv2d-51 [-1, 128, 56, 56] 32,768BatchNorm2d-52 [-1, 128, 56, 56] 256ReLU-53 [-1, 128, 56, 56] 0Conv2d-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] 144Conv2d-85 [-1, 4, 56, 56] 144BatchNorm2d-86 [-1, 128, 56, 56] 256ReLU-87 [-1, 128, 56, 56] 0GroupedConvBlock-88 [-1, 128, 56, 56] 0Conv2d-89 [-1, 256, 56, 56] 32,768BatchNorm2d-90 [-1, 256, 56, 56] 512ReLU-91 [-1, 256, 56, 56] 0ResNeXtBlock-92 [-1, 256, 56, 56] 0Conv2d-93 [-1, 256, 56, 56] 65,536BatchNorm2d-94 [-1, 256, 56, 56] 512Conv2d-95 [-1, 128, 56, 56] 32,768BatchNorm2d-96 [-1, 128, 56, 56] 256ReLU-97 [-1, 128, 56, 56] 0Conv2d-98 [-1, 4, 56, 56] 144Conv2d-99 [-1, 4, 56, 56] 144Conv2d-100 [-1, 4, 56, 56] 144Conv2d-101 [-1, 4, 56, 56] 144Conv2d-102 [-1, 4, 56, 56] 144Conv2d-103 [-1, 4, 56, 56] 144Conv2d-104 [-1, 4, 56, 56] 144Conv2d-105 [-1, 4, 56, 56] 144Conv2d-106 [-1, 4, 56, 56] 144Conv2d-107 [-1, 4, 56, 56] 144Conv2d-108 [-1, 4, 56, 56] 144Conv2d-109 [-1, 4, 56, 56] 144Conv2d-110 [-1, 4, 56, 56] 144Conv2d-111 [-1, 4, 56, 56] 144Conv2d-112 [-1, 4, 56, 56] 144Conv2d-113 [-1, 4, 56, 56] 144Conv2d-114 [-1, 4, 56, 56] 144Conv2d-115 [-1, 4, 56, 56] 144Conv2d-116 [-1, 4, 56, 56] 144Conv2d-117 [-1, 4, 56, 56] 144Conv2d-118 [-1, 4, 56, 56] 144Conv2d-119 [-1, 4, 56, 56] 144Conv2d-120 [-1, 4, 56, 56] 144Conv2d-121 [-1, 4, 56, 56] 144Conv2d-122 [-1, 4, 56, 56] 144Conv2d-123 [-1, 4, 56, 56] 144Conv2d-124 [-1, 4, 56, 56] 144Conv2d-125 [-1, 4, 56, 56] 144Conv2d-126 [-1, 4, 56, 56] 144Conv2d-127 [-1, 4, 56, 56] 144Conv2d-128 [-1, 4, 56, 56] 144Conv2d-129 [-1, 4, 56, 56] 144BatchNorm2d-130 [-1, 128, 56, 56] 256ReLU-131 [-1, 128, 56, 56] 0
GroupedConvBlock-132 [-1, 128, 56, 56] 0Conv2d-133 [-1, 256, 56, 56] 32,768BatchNorm2d-134 [-1, 256, 56, 56] 512ReLU-135 [-1, 256, 56, 56] 0ResNeXtBlock-136 [-1, 256, 56, 56] 0Conv2d-137 [-1, 512, 28, 28] 131,072BatchNorm2d-138 [-1, 512, 28, 28] 1,024Conv2d-139 [-1, 256, 56, 56] 65,536BatchNorm2d-140 [-1, 256, 56, 56] 512ReLU-141 [-1, 256, 56, 56] 0Conv2d-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] 576Conv2d-173 [-1, 8, 28, 28] 576BatchNorm2d-174 [-1, 256, 28, 28] 512ReLU-175 [-1, 256, 28, 28] 0
GroupedConvBlock-176 [-1, 256, 28, 28] 0Conv2d-177 [-1, 512, 28, 28] 131,072BatchNorm2d-178 [-1, 512, 28, 28] 1,024ReLU-179 [-1, 512, 28, 28] 0ResNeXtBlock-180 [-1, 512, 28, 28] 0Conv2d-181 [-1, 512, 28, 28] 262,144BatchNorm2d-182 [-1, 512, 28, 28] 1,024Conv2d-183 [-1, 256, 28, 28] 131,072BatchNorm2d-184 [-1, 256, 28, 28] 512ReLU-185 [-1, 256, 28, 28] 0Conv2d-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] 576Conv2d-216 [-1, 8, 28, 28] 576Conv2d-217 [-1, 8, 28, 28] 576BatchNorm2d-218 [-1, 256, 28, 28] 512ReLU-219 [-1, 256, 28, 28] 0
GroupedConvBlock-220 [-1, 256, 28, 28] 0Conv2d-221 [-1, 512, 28, 28] 131,072BatchNorm2d-222 [-1, 512, 28, 28] 1,024ReLU-223 [-1, 512, 28, 28] 0ResNeXtBlock-224 [-1, 512, 28, 28] 0Conv2d-225 [-1, 512, 28, 28] 262,144BatchNorm2d-226 [-1, 512, 28, 28] 1,024Conv2d-227 [-1, 256, 28, 28] 131,072BatchNorm2d-228 [-1, 256, 28, 28] 512ReLU-229 [-1, 256, 28, 28] 0Conv2d-230 [-1, 8, 28, 28] 576Conv2d-231 [-1, 8, 28, 28] 576Conv2d-232 [-1, 8, 28, 28] 576Conv2d-233 [-1, 8, 28, 28] 576Conv2d-234 [-1, 8, 28, 28] 576Conv2d-235 [-1, 8, 28, 28] 576Conv2d-236 [-1, 8, 28, 28] 576Conv2d-237 [-1, 8, 28, 28] 576Conv2d-238 [-1, 8, 28, 28] 576Conv2d-239 [-1, 8, 28, 28] 576Conv2d-240 [-1, 8, 28, 28] 576Conv2d-241 [-1, 8, 28, 28] 576Conv2d-242 [-1, 8, 28, 28] 576Conv2d-243 [-1, 8, 28, 28] 576Conv2d-244 [-1, 8, 28, 28] 576Conv2d-245 [-1, 8, 28, 28] 576Conv2d-246 [-1, 8, 28, 28] 576Conv2d-247 [-1, 8, 28, 28] 576Conv2d-248 [-1, 8, 28, 28] 576Conv2d-249 [-1, 8, 28, 28] 576Conv2d-250 [-1, 8, 28, 28] 576Conv2d-251 [-1, 8, 28, 28] 576Conv2d-252 [-1, 8, 28, 28] 576Conv2d-253 [-1, 8, 28, 28] 576Conv2d-254 [-1, 8, 28, 28] 576Conv2d-255 [-1, 8, 28, 28] 576Conv2d-256 [-1, 8, 28, 28] 576Conv2d-257 [-1, 8, 28, 28] 576Conv2d-258 [-1, 8, 28, 28] 576Conv2d-259 [-1, 8, 28, 28] 576Conv2d-260 [-1, 8, 28, 28] 576Conv2d-261 [-1, 8, 28, 28] 576BatchNorm2d-262 [-1, 256, 28, 28] 512ReLU-263 [-1, 256, 28, 28] 0
GroupedConvBlock-264 [-1, 256, 28, 28] 0Conv2d-265 [-1, 512, 28, 28] 131,072BatchNorm2d-266 [-1, 512, 28, 28] 1,024ReLU-267 [-1, 512, 28, 28] 0ResNeXtBlock-268 [-1, 512, 28, 28] 0Conv2d-269 [-1, 512, 28, 28] 262,144BatchNorm2d-270 [-1, 512, 28, 28] 1,024Conv2d-271 [-1, 256, 28, 28] 131,072BatchNorm2d-272 [-1, 256, 28, 28] 512ReLU-273 [-1, 256, 28, 28] 0Conv2d-274 [-1, 8, 28, 28] 576Conv2d-275 [-1, 8, 28, 28] 576Conv2d-276 [-1, 8, 28, 28] 576Conv2d-277 [-1, 8, 28, 28] 576Conv2d-278 [-1, 8, 28, 28] 576Conv2d-279 [-1, 8, 28, 28] 576Conv2d-280 [-1, 8, 28, 28] 576Conv2d-281 [-1, 8, 28, 28] 576Conv2d-282 [-1, 8, 28, 28] 576Conv2d-283 [-1, 8, 28, 28] 576Conv2d-284 [-1, 8, 28, 28] 576Conv2d-285 [-1, 8, 28, 28] 576Conv2d-286 [-1, 8, 28, 28] 576Conv2d-287 [-1, 8, 28, 28] 576Conv2d-288 [-1, 8, 28, 28] 576Conv2d-289 [-1, 8, 28, 28] 576Conv2d-290 [-1, 8, 28, 28] 576Conv2d-291 [-1, 8, 28, 28] 576Conv2d-292 [-1, 8, 28, 28] 576Conv2d-293 [-1, 8, 28, 28] 576Conv2d-294 [-1, 8, 28, 28] 576Conv2d-295 [-1, 8, 28, 28] 576Conv2d-296 [-1, 8, 28, 28] 576Conv2d-297 [-1, 8, 28, 28] 576Conv2d-298 [-1, 8, 28, 28] 576Conv2d-299 [-1, 8, 28, 28] 576Conv2d-300 [-1, 8, 28, 28] 576Conv2d-301 [-1, 8, 28, 28] 576Conv2d-302 [-1, 8, 28, 28] 576Conv2d-303 [-1, 8, 28, 28] 576Conv2d-304 [-1, 8, 28, 28] 576Conv2d-305 [-1, 8, 28, 28] 576BatchNorm2d-306 [-1, 256, 28, 28] 512ReLU-307 [-1, 256, 28, 28] 0
GroupedConvBlock-308 [-1, 256, 28, 28] 0Conv2d-309 [-1, 512, 28, 28] 131,072BatchNorm2d-310 [-1, 512, 28, 28] 1,024ReLU-311 [-1, 512, 28, 28] 0ResNeXtBlock-312 [-1, 512, 28, 28] 0Conv2d-313 [-1, 1024, 14, 14] 524,288BatchNorm2d-314 [-1, 1024, 14, 14] 2,048Conv2d-315 [-1, 512, 28, 28] 262,144BatchNorm2d-316 [-1, 512, 28, 28] 1,024ReLU-317 [-1, 512, 28, 28] 0Conv2d-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] 2,304Conv2d-347 [-1, 16, 14, 14] 2,304Conv2d-348 [-1, 16, 14, 14] 2,304Conv2d-349 [-1, 16, 14, 14] 2,304BatchNorm2d-350 [-1, 512, 14, 14] 1,024ReLU-351 [-1, 512, 14, 14] 0
GroupedConvBlock-352 [-1, 512, 14, 14] 0Conv2d-353 [-1, 1024, 14, 14] 524,288BatchNorm2d-354 [-1, 1024, 14, 14] 2,048ReLU-355 [-1, 1024, 14, 14] 0ResNeXtBlock-356 [-1, 1024, 14, 14] 0Conv2d-357 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-358 [-1, 1024, 14, 14] 2,048Conv2d-359 [-1, 512, 14, 14] 524,288BatchNorm2d-360 [-1, 512, 14, 14] 1,024ReLU-361 [-1, 512, 14, 14] 0Conv2d-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] 2,304Conv2d-390 [-1, 16, 14, 14] 2,304Conv2d-391 [-1, 16, 14, 14] 2,304Conv2d-392 [-1, 16, 14, 14] 2,304Conv2d-393 [-1, 16, 14, 14] 2,304BatchNorm2d-394 [-1, 512, 14, 14] 1,024ReLU-395 [-1, 512, 14, 14] 0
GroupedConvBlock-396 [-1, 512, 14, 14] 0Conv2d-397 [-1, 1024, 14, 14] 524,288BatchNorm2d-398 [-1, 1024, 14, 14] 2,048ReLU-399 [-1, 1024, 14, 14] 0ResNeXtBlock-400 [-1, 1024, 14, 14] 0Conv2d-401 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-402 [-1, 1024, 14, 14] 2,048Conv2d-403 [-1, 512, 14, 14] 524,288BatchNorm2d-404 [-1, 512, 14, 14] 1,024ReLU-405 [-1, 512, 14, 14] 0Conv2d-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] 2,304Conv2d-433 [-1, 16, 14, 14] 2,304Conv2d-434 [-1, 16, 14, 14] 2,304Conv2d-435 [-1, 16, 14, 14] 2,304Conv2d-436 [-1, 16, 14, 14] 2,304Conv2d-437 [-1, 16, 14, 14] 2,304BatchNorm2d-438 [-1, 512, 14, 14] 1,024ReLU-439 [-1, 512, 14, 14] 0
GroupedConvBlock-440 [-1, 512, 14, 14] 0Conv2d-441 [-1, 1024, 14, 14] 524,288BatchNorm2d-442 [-1, 1024, 14, 14] 2,048ReLU-443 [-1, 1024, 14, 14] 0ResNeXtBlock-444 [-1, 1024, 14, 14] 0Conv2d-445 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-446 [-1, 1024, 14, 14] 2,048Conv2d-447 [-1, 512, 14, 14] 524,288BatchNorm2d-448 [-1, 512, 14, 14] 1,024ReLU-449 [-1, 512, 14, 14] 0Conv2d-450 [-1, 16, 14, 14] 2,304Conv2d-451 [-1, 16, 14, 14] 2,304Conv2d-452 [-1, 16, 14, 14] 2,304Conv2d-453 [-1, 16, 14, 14] 2,304Conv2d-454 [-1, 16, 14, 14] 2,304Conv2d-455 [-1, 16, 14, 14] 2,304Conv2d-456 [-1, 16, 14, 14] 2,304Conv2d-457 [-1, 16, 14, 14] 2,304Conv2d-458 [-1, 16, 14, 14] 2,304Conv2d-459 [-1, 16, 14, 14] 2,304Conv2d-460 [-1, 16, 14, 14] 2,304Conv2d-461 [-1, 16, 14, 14] 2,304Conv2d-462 [-1, 16, 14, 14] 2,304Conv2d-463 [-1, 16, 14, 14] 2,304Conv2d-464 [-1, 16, 14, 14] 2,304Conv2d-465 [-1, 16, 14, 14] 2,304Conv2d-466 [-1, 16, 14, 14] 2,304Conv2d-467 [-1, 16, 14, 14] 2,304Conv2d-468 [-1, 16, 14, 14] 2,304Conv2d-469 [-1, 16, 14, 14] 2,304Conv2d-470 [-1, 16, 14, 14] 2,304Conv2d-471 [-1, 16, 14, 14] 2,304Conv2d-472 [-1, 16, 14, 14] 2,304Conv2d-473 [-1, 16, 14, 14] 2,304Conv2d-474 [-1, 16, 14, 14] 2,304Conv2d-475 [-1, 16, 14, 14] 2,304Conv2d-476 [-1, 16, 14, 14] 2,304Conv2d-477 [-1, 16, 14, 14] 2,304Conv2d-478 [-1, 16, 14, 14] 2,304Conv2d-479 [-1, 16, 14, 14] 2,304Conv2d-480 [-1, 16, 14, 14] 2,304Conv2d-481 [-1, 16, 14, 14] 2,304BatchNorm2d-482 [-1, 512, 14, 14] 1,024ReLU-483 [-1, 512, 14, 14] 0
GroupedConvBlock-484 [-1, 512, 14, 14] 0Conv2d-485 [-1, 1024, 14, 14] 524,288BatchNorm2d-486 [-1, 1024, 14, 14] 2,048ReLU-487 [-1, 1024, 14, 14] 0ResNeXtBlock-488 [-1, 1024, 14, 14] 0Conv2d-489 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-490 [-1, 1024, 14, 14] 2,048Conv2d-491 [-1, 512, 14, 14] 524,288BatchNorm2d-492 [-1, 512, 14, 14] 1,024ReLU-493 [-1, 512, 14, 14] 0Conv2d-494 [-1, 16, 14, 14] 2,304Conv2d-495 [-1, 16, 14, 14] 2,304Conv2d-496 [-1, 16, 14, 14] 2,304Conv2d-497 [-1, 16, 14, 14] 2,304Conv2d-498 [-1, 16, 14, 14] 2,304Conv2d-499 [-1, 16, 14, 14] 2,304Conv2d-500 [-1, 16, 14, 14] 2,304Conv2d-501 [-1, 16, 14, 14] 2,304Conv2d-502 [-1, 16, 14, 14] 2,304Conv2d-503 [-1, 16, 14, 14] 2,304Conv2d-504 [-1, 16, 14, 14] 2,304Conv2d-505 [-1, 16, 14, 14] 2,304Conv2d-506 [-1, 16, 14, 14] 2,304Conv2d-507 [-1, 16, 14, 14] 2,304Conv2d-508 [-1, 16, 14, 14] 2,304Conv2d-509 [-1, 16, 14, 14] 2,304Conv2d-510 [-1, 16, 14, 14] 2,304Conv2d-511 [-1, 16, 14, 14] 2,304Conv2d-512 [-1, 16, 14, 14] 2,304Conv2d-513 [-1, 16, 14, 14] 2,304Conv2d-514 [-1, 16, 14, 14] 2,304Conv2d-515 [-1, 16, 14, 14] 2,304Conv2d-516 [-1, 16, 14, 14] 2,304Conv2d-517 [-1, 16, 14, 14] 2,304Conv2d-518 [-1, 16, 14, 14] 2,304Conv2d-519 [-1, 16, 14, 14] 2,304Conv2d-520 [-1, 16, 14, 14] 2,304Conv2d-521 [-1, 16, 14, 14] 2,304Conv2d-522 [-1, 16, 14, 14] 2,304Conv2d-523 [-1, 16, 14, 14] 2,304Conv2d-524 [-1, 16, 14, 14] 2,304Conv2d-525 [-1, 16, 14, 14] 2,304BatchNorm2d-526 [-1, 512, 14, 14] 1,024ReLU-527 [-1, 512, 14, 14] 0
GroupedConvBlock-528 [-1, 512, 14, 14] 0Conv2d-529 [-1, 1024, 14, 14] 524,288BatchNorm2d-530 [-1, 1024, 14, 14] 2,048ReLU-531 [-1, 1024, 14, 14] 0ResNeXtBlock-532 [-1, 1024, 14, 14] 0Conv2d-533 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-534 [-1, 1024, 14, 14] 2,048Conv2d-535 [-1, 512, 14, 14] 524,288BatchNorm2d-536 [-1, 512, 14, 14] 1,024ReLU-537 [-1, 512, 14, 14] 0Conv2d-538 [-1, 16, 14, 14] 2,304Conv2d-539 [-1, 16, 14, 14] 2,304Conv2d-540 [-1, 16, 14, 14] 2,304Conv2d-541 [-1, 16, 14, 14] 2,304Conv2d-542 [-1, 16, 14, 14] 2,304Conv2d-543 [-1, 16, 14, 14] 2,304Conv2d-544 [-1, 16, 14, 14] 2,304Conv2d-545 [-1, 16, 14, 14] 2,304Conv2d-546 [-1, 16, 14, 14] 2,304Conv2d-547 [-1, 16, 14, 14] 2,304Conv2d-548 [-1, 16, 14, 14] 2,304Conv2d-549 [-1, 16, 14, 14] 2,304Conv2d-550 [-1, 16, 14, 14] 2,304Conv2d-551 [-1, 16, 14, 14] 2,304Conv2d-552 [-1, 16, 14, 14] 2,304Conv2d-553 [-1, 16, 14, 14] 2,304Conv2d-554 [-1, 16, 14, 14] 2,304Conv2d-555 [-1, 16, 14, 14] 2,304Conv2d-556 [-1, 16, 14, 14] 2,304Conv2d-557 [-1, 16, 14, 14] 2,304Conv2d-558 [-1, 16, 14, 14] 2,304Conv2d-559 [-1, 16, 14, 14] 2,304Conv2d-560 [-1, 16, 14, 14] 2,304Conv2d-561 [-1, 16, 14, 14] 2,304Conv2d-562 [-1, 16, 14, 14] 2,304Conv2d-563 [-1, 16, 14, 14] 2,304Conv2d-564 [-1, 16, 14, 14] 2,304Conv2d-565 [-1, 16, 14, 14] 2,304Conv2d-566 [-1, 16, 14, 14] 2,304Conv2d-567 [-1, 16, 14, 14] 2,304Conv2d-568 [-1, 16, 14, 14] 2,304Conv2d-569 [-1, 16, 14, 14] 2,304BatchNorm2d-570 [-1, 512, 14, 14] 1,024ReLU-571 [-1, 512, 14, 14] 0
GroupedConvBlock-572 [-1, 512, 14, 14] 0Conv2d-573 [-1, 1024, 14, 14] 524,288BatchNorm2d-574 [-1, 1024, 14, 14] 2,048ReLU-575 [-1, 1024, 14, 14] 0ResNeXtBlock-576 [-1, 1024, 14, 14] 0Conv2d-577 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-578 [-1, 2048, 7, 7] 4,096Conv2d-579 [-1, 1024, 14, 14] 1,048,576BatchNorm2d-580 [-1, 1024, 14, 14] 2,048ReLU-581 [-1, 1024, 14, 14] 0Conv2d-582 [-1, 32, 7, 7] 9,216Conv2d-583 [-1, 32, 7, 7] 9,216Conv2d-584 [-1, 32, 7, 7] 9,216Conv2d-585 [-1, 32, 7, 7] 9,216Conv2d-586 [-1, 32, 7, 7] 9,216Conv2d-587 [-1, 32, 7, 7] 9,216Conv2d-588 [-1, 32, 7, 7] 9,216Conv2d-589 [-1, 32, 7, 7] 9,216Conv2d-590 [-1, 32, 7, 7] 9,216Conv2d-591 [-1, 32, 7, 7] 9,216Conv2d-592 [-1, 32, 7, 7] 9,216Conv2d-593 [-1, 32, 7, 7] 9,216Conv2d-594 [-1, 32, 7, 7] 9,216Conv2d-595 [-1, 32, 7, 7] 9,216Conv2d-596 [-1, 32, 7, 7] 9,216Conv2d-597 [-1, 32, 7, 7] 9,216Conv2d-598 [-1, 32, 7, 7] 9,216Conv2d-599 [-1, 32, 7, 7] 9,216Conv2d-600 [-1, 32, 7, 7] 9,216Conv2d-601 [-1, 32, 7, 7] 9,216Conv2d-602 [-1, 32, 7, 7] 9,216Conv2d-603 [-1, 32, 7, 7] 9,216Conv2d-604 [-1, 32, 7, 7] 9,216Conv2d-605 [-1, 32, 7, 7] 9,216Conv2d-606 [-1, 32, 7, 7] 9,216Conv2d-607 [-1, 32, 7, 7] 9,216Conv2d-608 [-1, 32, 7, 7] 9,216Conv2d-609 [-1, 32, 7, 7] 9,216Conv2d-610 [-1, 32, 7, 7] 9,216Conv2d-611 [-1, 32, 7, 7] 9,216Conv2d-612 [-1, 32, 7, 7] 9,216Conv2d-613 [-1, 32, 7, 7] 9,216BatchNorm2d-614 [-1, 1024, 7, 7] 2,048ReLU-615 [-1, 1024, 7, 7] 0
GroupedConvBlock-616 [-1, 1024, 7, 7] 0Conv2d-617 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-618 [-1, 2048, 7, 7] 4,096ReLU-619 [-1, 2048, 7, 7] 0ResNeXtBlock-620 [-1, 2048, 7, 7] 0Conv2d-621 [-1, 2048, 7, 7] 4,194,304BatchNorm2d-622 [-1, 2048, 7, 7] 4,096Conv2d-623 [-1, 1024, 7, 7] 2,097,152BatchNorm2d-624 [-1, 1024, 7, 7] 2,048ReLU-625 [-1, 1024, 7, 7] 0Conv2d-626 [-1, 32, 7, 7] 9,216Conv2d-627 [-1, 32, 7, 7] 9,216Conv2d-628 [-1, 32, 7, 7] 9,216Conv2d-629 [-1, 32, 7, 7] 9,216Conv2d-630 [-1, 32, 7, 7] 9,216Conv2d-631 [-1, 32, 7, 7] 9,216Conv2d-632 [-1, 32, 7, 7] 9,216Conv2d-633 [-1, 32, 7, 7] 9,216Conv2d-634 [-1, 32, 7, 7] 9,216Conv2d-635 [-1, 32, 7, 7] 9,216Conv2d-636 [-1, 32, 7, 7] 9,216Conv2d-637 [-1, 32, 7, 7] 9,216Conv2d-638 [-1, 32, 7, 7] 9,216Conv2d-639 [-1, 32, 7, 7] 9,216Conv2d-640 [-1, 32, 7, 7] 9,216Conv2d-641 [-1, 32, 7, 7] 9,216Conv2d-642 [-1, 32, 7, 7] 9,216Conv2d-643 [-1, 32, 7, 7] 9,216Conv2d-644 [-1, 32, 7, 7] 9,216Conv2d-645 [-1, 32, 7, 7] 9,216Conv2d-646 [-1, 32, 7, 7] 9,216Conv2d-647 [-1, 32, 7, 7] 9,216Conv2d-648 [-1, 32, 7, 7] 9,216Conv2d-649 [-1, 32, 7, 7] 9,216Conv2d-650 [-1, 32, 7, 7] 9,216Conv2d-651 [-1, 32, 7, 7] 9,216Conv2d-652 [-1, 32, 7, 7] 9,216Conv2d-653 [-1, 32, 7, 7] 9,216Conv2d-654 [-1, 32, 7, 7] 9,216Conv2d-655 [-1, 32, 7, 7] 9,216Conv2d-656 [-1, 32, 7, 7] 9,216Conv2d-657 [-1, 32, 7, 7] 9,216BatchNorm2d-658 [-1, 1024, 7, 7] 2,048ReLU-659 [-1, 1024, 7, 7] 0
GroupedConvBlock-660 [-1, 1024, 7, 7] 0Conv2d-661 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-662 [-1, 2048, 7, 7] 4,096ReLU-663 [-1, 2048, 7, 7] 0ResNeXtBlock-664 [-1, 2048, 7, 7] 0Conv2d-665 [-1, 2048, 7, 7] 4,194,304BatchNorm2d-666 [-1, 2048, 7, 7] 4,096Conv2d-667 [-1, 1024, 7, 7] 2,097,152BatchNorm2d-668 [-1, 1024, 7, 7] 2,048ReLU-669 [-1, 1024, 7, 7] 0Conv2d-670 [-1, 32, 7, 7] 9,216Conv2d-671 [-1, 32, 7, 7] 9,216Conv2d-672 [-1, 32, 7, 7] 9,216Conv2d-673 [-1, 32, 7, 7] 9,216Conv2d-674 [-1, 32, 7, 7] 9,216Conv2d-675 [-1, 32, 7, 7] 9,216Conv2d-676 [-1, 32, 7, 7] 9,216Conv2d-677 [-1, 32, 7, 7] 9,216Conv2d-678 [-1, 32, 7, 7] 9,216Conv2d-679 [-1, 32, 7, 7] 9,216Conv2d-680 [-1, 32, 7, 7] 9,216Conv2d-681 [-1, 32, 7, 7] 9,216Conv2d-682 [-1, 32, 7, 7] 9,216Conv2d-683 [-1, 32, 7, 7] 9,216Conv2d-684 [-1, 32, 7, 7] 9,216Conv2d-685 [-1, 32, 7, 7] 9,216Conv2d-686 [-1, 32, 7, 7] 9,216Conv2d-687 [-1, 32, 7, 7] 9,216Conv2d-688 [-1, 32, 7, 7] 9,216Conv2d-689 [-1, 32, 7, 7] 9,216Conv2d-690 [-1, 32, 7, 7] 9,216Conv2d-691 [-1, 32, 7, 7] 9,216Conv2d-692 [-1, 32, 7, 7] 9,216Conv2d-693 [-1, 32, 7, 7] 9,216Conv2d-694 [-1, 32, 7, 7] 9,216Conv2d-695 [-1, 32, 7, 7] 9,216Conv2d-696 [-1, 32, 7, 7] 9,216Conv2d-697 [-1, 32, 7, 7] 9,216Conv2d-698 [-1, 32, 7, 7] 9,216Conv2d-699 [-1, 32, 7, 7] 9,216Conv2d-700 [-1, 32, 7, 7] 9,216Conv2d-701 [-1, 32, 7, 7] 9,216BatchNorm2d-702 [-1, 1024, 7, 7] 2,048ReLU-703 [-1, 1024, 7, 7] 0
GroupedConvBlock-704 [-1, 1024, 7, 7] 0Conv2d-705 [-1, 2048, 7, 7] 2,097,152BatchNorm2d-706 [-1, 2048, 7, 7] 4,096ReLU-707 [-1, 2048, 7, 7] 0ResNeXtBlock-708 [-1, 2048, 7, 7] 0
AdaptiveAvgPool2d-709 [-1, 2048, 1, 1] 0Linear-710 [-1, 4] 8,196
================================================================
Total params: 37,559,620
Trainable params: 37,559,620
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 444.08
Params size (MB): 143.28
Estimated Total Size (MB): 587.93
----------------------------------------------------------------
# 定义损失函数和优化器
loss_fn = nn.CrossEntropyLoss()learn_rate = 5e-4
lambda1 = lambda epc: (0.92 ** (epc // 2))optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) # 选定调整方法# 训练函数
def train(dataloader,model,loss_fn,optimizer):size = len(dataloader.dataset) # 训练集大小num_batches = len(dataloader) # 批次数目train_loss,train_acc = 0,0for X,y in dataloader:X,y = X.to(device),y.to(device)# 计算预测误差pred = model(X)loss = loss_fn(pred,y)# 反向传播optimizer.zero_grad()loss.backward()optimizer.step()# 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_loss# 测试函数
def test(dataloader,model,loss_fn):size = len(dataloader.dataset)num_batches = len(dataloader)test_acc,test_loss = 0,0with torch.no_grad():for X,y in dataloader:X,y = X.to(device),y.to(device)# 计算losspred = model(X)loss = loss_fn(pred,y)test_acc += (pred.argmax(1)==y).type(torch.float).sum().item()test_loss += loss.item()test_acc /= sizetest_loss /= num_batchesreturn test_acc,test_lossimport copyepochs = 20train_acc = []
train_loss = []
test_acc = []
test_loss = []best_acc = 0.0# 检查 GPU 可用性并打印设备信息
if torch.cuda.is_available():for i in range(torch.cuda.device_count()):print(f"GPU {i}: {torch.cuda.get_device_name(i)}")print(f"Initial Memory Allocated: {torch.cuda.memory_allocated(i)/1024**2:.2f} MB")print(f"Initial Memory Cached: {torch.cuda.memory_reserved(i)/1024**2:.2f} MB")
else:print("No GPU available. Using CPU.")# 多显卡设置 当前使用的是使用 PyTorch 自带的 DataParallel，后续如有需要可以设置为DistributedDataParallel，这是更加高效的方式
# 且多卡不一定比单卡效果就好，需要调整优化
# if torch.cuda.device_count() > 1:
# print(f"Using {torch.cuda.device_count()} GPUs")
# model = nn.DataParallel(model)
# model = model.to('cuda')for epoch in range(epochs):# 更新学习率——使用自定义学习率时使用# adjust_learning_rate(optimizer,epoch,learn_rate)model.train()epoch_train_acc,epoch_train_loss = train(train_dl,model,loss_fn,optimizer)scheduler.step() # 更新学习率——调用官方动态学习率时使用model.eval()epoch_test_acc,epoch_test_loss = test(test_dl,model,loss_fn)# 保存最佳模型到 best_modelif epoch_test_acc > best_acc:best_acc = epoch_test_accbest_model = copy.deepcopy(model)train_acc.append(epoch_train_acc)train_loss.append(epoch_train_loss)test_acc.append(epoch_test_acc)test_loss.append(epoch_test_loss)# 获取当前学习率lr = optimizer.state_dict()['param_groups'][0]['lr']template = ('Epoch:{:2d},Train_acc:{:.1f}%,Train_loss:{:.3f},Test_acc:{:.1f}%,Test_loss:{:.3f},Lr:{:.2E}')print(template.format(epoch+1,epoch_train_acc*100,epoch_train_loss,epoch_test_acc*100,epoch_test_loss,lr))# 实时监控 GPU 状态if torch.cuda.is_available():for i in range(torch.cuda.device_count()):print(f"GPU {i} Usage:")print(f" Memory Allocated: {torch.cuda.memory_allocated(i)/1024**2:.2f} MB")print(f" Memory Cached: {torch.cuda.memory_reserved(i)/1024**2:.2f} MB")print(f" Max Memory Allocated: {torch.cuda.max_memory_allocated(i)/1024**2:.2f} MB")print(f" Max Memory Cached: {torch.cuda.max_memory_reserved(i)/1024**2:.2f} MB")print('Done','best_acc: ',best_acc)
GPU 0: NVIDIA GeForce RTX 2070
Initial Memory Allocated: 151.86 MB
Initial Memory Cached: 606.00 MB
Epoch: 1,Train_acc:36.5%,Train_loss:1.495,Test_acc:48.7%,Test_loss:1.233,Lr:5.00E-04
GPU 0 Usage:Memory Allocated: 743.27 MBMemory Cached: 2194.00 MBMax Memory Allocated: 1733.31 MBMax Memory Cached: 2194.00 MB
Epoch: 2,Train_acc:50.4%,Train_loss:1.141,Test_acc:58.4%,Test_loss:0.928,Lr:4.60E-04
GPU 0 Usage:Memory Allocated: 738.20 MBMemory Cached: 2208.00 MBMax Memory Allocated: 1882.58 MBMax Memory Cached: 2208.00 MB
Epoch: 3,Train_acc:56.4%,Train_loss:1.019,Test_acc:61.9%,Test_loss:0.943,Lr:4.60E-04
GPU 0 Usage:Memory Allocated: 737.86 MBMemory Cached: 2208.00 MBMax Memory Allocated: 1883.39 MBMax Memory Cached: 2208.00 MB
Epoch: 4,Train_acc:65.9%,Train_loss:0.851,Test_acc:63.7%,Test_loss:0.890,Lr:4.23E-04
GPU 0 Usage:Memory Allocated: 737.74 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1883.39 MBMax Memory Cached: 2210.00 MB
Epoch: 5,Train_acc:72.3%,Train_loss:0.746,Test_acc:77.9%,Test_loss:0.636,Lr:4.23E-04
GPU 0 Usage:Memory Allocated: 738.58 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.70 MBMax Memory Cached: 2210.00 MB
Epoch: 6,Train_acc:71.9%,Train_loss:0.770,Test_acc:79.6%,Test_loss:0.512,Lr:3.89E-04
GPU 0 Usage:Memory Allocated: 737.86 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.70 MBMax Memory Cached: 2210.00 MB
Epoch: 7,Train_acc:73.2%,Train_loss:0.677,Test_acc:80.5%,Test_loss:0.643,Lr:3.89E-04
GPU 0 Usage:Memory Allocated: 737.42 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.70 MBMax Memory Cached: 2210.00 MB
Epoch: 8,Train_acc:71.9%,Train_loss:0.709,Test_acc:86.7%,Test_loss:0.476,Lr:3.58E-04
GPU 0 Usage:Memory Allocated: 737.61 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.70 MBMax Memory Cached: 2210.00 MB
Epoch: 9,Train_acc:73.7%,Train_loss:0.658,Test_acc:70.8%,Test_loss:0.712,Lr:3.58E-04
GPU 0 Usage:Memory Allocated: 737.95 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.70 MBMax Memory Cached: 2210.00 MB
Epoch:10,Train_acc:77.4%,Train_loss:0.645,Test_acc:85.0%,Test_loss:0.414,Lr:3.30E-04
GPU 0 Usage:Memory Allocated: 737.95 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.89 MBMax Memory Cached: 2210.00 MB
Epoch:11,Train_acc:77.7%,Train_loss:0.654,Test_acc:83.2%,Test_loss:0.527,Lr:3.30E-04
GPU 0 Usage:Memory Allocated: 737.27 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1884.89 MBMax Memory Cached: 2210.00 MB
Epoch:12,Train_acc:81.0%,Train_loss:0.561,Test_acc:87.6%,Test_loss:0.403,Lr:3.03E-04
GPU 0 Usage:Memory Allocated: 736.58 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:13,Train_acc:81.9%,Train_loss:0.482,Test_acc:81.4%,Test_loss:0.472,Lr:3.03E-04
GPU 0 Usage:Memory Allocated: 736.74 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:14,Train_acc:81.9%,Train_loss:0.539,Test_acc:85.0%,Test_loss:0.380,Lr:2.79E-04
GPU 0 Usage:Memory Allocated: 737.14 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:15,Train_acc:83.0%,Train_loss:0.496,Test_acc:86.7%,Test_loss:0.369,Lr:2.79E-04
GPU 0 Usage:Memory Allocated: 741.05 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:16,Train_acc:80.8%,Train_loss:0.488,Test_acc:86.7%,Test_loss:0.449,Lr:2.57E-04
GPU 0 Usage:Memory Allocated: 741.00 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:17,Train_acc:82.5%,Train_loss:0.477,Test_acc:83.2%,Test_loss:0.471,Lr:2.57E-04
GPU 0 Usage:Memory Allocated: 741.89 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:18,Train_acc:86.5%,Train_loss:0.422,Test_acc:84.1%,Test_loss:0.513,Lr:2.36E-04
GPU 0 Usage:Memory Allocated: 741.00 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:19,Train_acc:81.6%,Train_loss:0.494,Test_acc:82.3%,Test_loss:0.414,Lr:2.36E-04
GPU 0 Usage:Memory Allocated: 741.89 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Epoch:20,Train_acc:83.8%,Train_loss:0.423,Test_acc:91.2%,Test_loss:0.276,Lr:2.17E-04
GPU 0 Usage:Memory Allocated: 741.30 MBMemory Cached: 2210.00 MBMax Memory Allocated: 1886.42 MBMax Memory Cached: 2210.00 MB
Done best_acc: 0.911504424778761
# 结果可视化
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.subplot(1,2,2)
plt.plot(epochs_range,train_loss,label = 'Test Accuracy')
plt.plot(epochs_range,test_loss,label = 'Test Loss')
plt.legend(loc = 'lower right')
plt.title('Training and validation Loss')
plt.show()# 自定义模型保存
# 状态字典保存
torch.save(model.state_dict(),r'C:\Users\11054\Desktop\kLearning\J6_learning\J6_ResNeXt50.pth') # 仅保存状态字典# 定义模型用来加载参数
best_model = ResNeXt50(num_classes=len(class_names)).to(device)best_model.load_state_dict(torch.load(r'C:\Users\11054\Desktop\kLearning\J6_learning\J6_ResNeXt50.pth')) # 加载状态字典到模型
<All keys matched successfully>
# 预测指定模型
from PIL import Image
import torchvision.transforms as transformsclasses = list(total_data.class_to_idx) # classes = list(total_data.class_to_idx)def predict_one_image(image_path,model,transform,classes):test_img = Image.open(image_path).convert('RGB')test_img = transform(test_img)img = test_img.to(device).unsqueeze(0)model.eval()output = model(img)print(output) # 观察模型预测结果的输出数据_,pred = torch.max(output,1)pred_class = classes[pred]print(f'预测结果是:{pred_class}')# 预测训练集中的某张照片
predict_one_image(image_path=r'C:\Users\11054\Desktop\kLearning\J1_learning\bird_photos\Bananaquit\001.jpg',model = model,transform = test_transforms,classes = classes)tensor([[ 1.1780, -1.3850, 1.1631, -1.2493]], device='cuda:0',grad_fn=<AddmmBackward0>)
预测结果是:Bananaquit
个人总结
ResNeXt-50特点
- 分组卷积（Grouped Convolutions）：
ResNeXt 引入了分组卷积的概念，将输入特征图分成多个组（group），每个组独立进行卷积操作，然后将结果拼接起来。
2. 残差连接（Residual Connections）：
与 ResNet 类似，ResNeXt 也使用了残差连接，允许信息在网络中直接跳过某些层，从而缓解梯度消失问题，并使得训练更深的网络变得可行。
ResNeXt-50 中的每个残差块（residual block）由多个分组卷积层组成，每个分组卷积层都包含 1x1、3x3、1x1 三个卷积核。
3. 网络结构：
ResNeXt-50 的网络结构主要由几个阶段组成，每个阶段包含多个残差块。典型的结构如下：
阶段 1：7x7 卷积层 + 3x3 最大池化层
阶段 2：包含 3 个残差块，每个残差块的输出通道数为 256
阶段 3：包含 4 个残差块，每个残差块的输出通道数为 512
阶段 4：包含 6 个残差块，每个残差块的输出通道数为 1024
阶段 5：包含 3 个残差块，每个残差块的输出通道数为 2048
最后是全局平均池化层和全连接层，用于分类任务。
4. 性能提升：
ResNeXt-50 通过增加基数（cardinality）而不显著增加模型参数的数量，实现了比 ResNet-50 更好的性能。
