PyTorch 中级篇（2）：深度残差网络（Deep Residual Networks）

发表于 更新于 分类于
Pytorch中级S02E02：深度残差网络（Deep Residual Networks）。
深度残差网络
残差单元
模型可视化工具Netron

PyTorch

PyTorch 中级篇（2）：深度残差网络（Deep Residual Networks）

参考代码

yunjey的 pytorch tutorial系列

深度残差网络 学习资源

论文原文

Deep Residual Learning for Image Recognition

Kaiming He的深度残差网络PPT

秒懂！何凯明的深度残差网络PPT是这样的|ICML2016 tutorial

Pytorch实现

根据原文【4.2. CIFAR-10 and Analysis】一节设计的针对数据集CIFAR-10的深度残差网络。

预处理

1
2
3
4
5
# 包
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# 设备配置
torch.cuda.set_device(1) # 这句用来设置pytorch在哪块GPU上运行
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# 超参数设置
num_epochs = 80
learning_rate = 0.001

# 图像预处理模块
# 先padding ，再 翻转，然后 裁剪。数据增广的手段
transform = transforms.Compose([
    transforms.Pad(4),
    transforms.RandomHorizontalFlip(),
    transforms.RandomCrop(32),
    transforms.ToTensor()])

CIFAR-10 数据集

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
#  训练数据集
train_dataset = torchvision.datasets.CIFAR10(root='../../../data/cifar-10',
                                             train=True, 
                                             transform=transform,
                                             download=True)

# 测试数据集
test_dataset = torchvision.datasets.CIFAR10(root='../../../data/cifar-10',
                                            train=False, 
                                            transform=transforms.ToTensor())

# 数据加载器
# 训练数据加载器
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
                                           batch_size=100, 
                                           shuffle=True)
# 测试数据加载器
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
                                          batch_size=100, 
                                          shuffle=False)
Files already downloaded and verified

深度残差网络模型设计

3x3卷积层

1
2
3
4
# 3x3 convolution
def conv3x3(in_channels, out_channels, stride=1):
    return nn.Conv2d(in_channels, out_channels, kernel_size=3, 
                     stride=stride, padding=1, bias=False)

残差块（残差单元）（Residual block）

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# Residual block
class ResidualBlock(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, downsample=None):
        super(ResidualBlock, self).__init__()
        self.conv1 = conv3x3(in_channels, out_channels, stride)
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(out_channels, out_channels)
        self.bn2 = nn.BatchNorm2d(out_channels)
        self.downsample = downsample
        
    def forward(self, x):
        residual = x
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        if self.downsample:
            residual = self.downsample(x)
        out += residual
        out = self.relu(out)
        return out

残差网络搭建

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
# ResNet
class ResNet(nn.Module):
    def __init__(self, block, layers, num_classes=10):
        super(ResNet, self).__init__()
        self.in_channels = 16
        self.conv = conv3x3(3, 16)
        self.bn = nn.BatchNorm2d(16)
        self.relu = nn.ReLU(inplace=True)
        self.layer1 = self.make_layer(block, 16, layers[0])
        self.layer2 = self.make_layer(block, 32, layers[0], 2)
        self.layer3 = self.make_layer(block, 64, layers[1], 2)
        self.avg_pool = nn.AvgPool2d(8,ceil_mode=False) #  nn.AvgPool2d需要添加参数ceil_mode=False，否则该模块无法导出为onnx格式
        self.fc = nn.Linear(64, num_classes)
        
    def make_layer(self, block, out_channels, blocks, stride=1):
        downsample = None
        if (stride != 1) or (self.in_channels != out_channels):
            downsample = nn.Sequential(
                conv3x3(self.in_channels, out_channels, stride=stride),
                nn.BatchNorm2d(out_channels))
        layers = []
        layers.append(block(self.in_channels, out_channels, stride, downsample)) # 残差直接映射部分
        self.in_channels = out_channels
        for i in range(1, blocks):
            layers.append(block(out_channels, out_channels))
        return nn.Sequential(*layers)
    
    def forward(self, x):
        out = self.conv(x)
        out = self.bn(out)
        out = self.relu(out)
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.avg_pool(out)
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        return out

实例化模型

1
2
3
4
5
6
7
8
9
10
11
# 实例化一个残差网络模型
model = ResNet(ResidualBlock, [2, 2, 2, 2]).to(device)

# 设置损失函数和优化器
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

# 用于更新参数组中的学习率
def update_lr(optimizer, lr):    
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

训练模型

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
total_step = len(train_loader)
curr_lr = learning_rate
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(train_loader):
        images = images.to(device)
        labels = labels.to(device)
        
        # 前向传播
        outputs = model(images)
        loss = criterion(outputs, labels)
        
        # 反向传播和优化
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if ((i+1) % 100 == 0) and ((epoch+1) % 5 == 0):
            print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}"
                   .format(epoch+1, num_epochs, i+1, total_step, loss.item()))

    # 学习率衰减策略
    if (epoch+1) % 20 == 0:
        curr_lr /= 3
        update_lr(optimizer, curr_lr)
Epoch [5/80], Step [100/500] Loss: 0.6474
Epoch [5/80], Step [200/500] Loss: 0.7043
Epoch [5/80], Step [300/500] Loss: 0.7472
Epoch [5/80], Step [400/500] Loss: 0.6662
Epoch [5/80], Step [500/500] Loss: 0.6378
Epoch [10/80], Step [100/500] Loss: 0.5786
Epoch [10/80], Step [200/500] Loss: 0.7229
Epoch [10/80], Step [300/500] Loss: 0.6183
Epoch [10/80], Step [400/500] Loss: 0.5043
Epoch [10/80], Step [500/500] Loss: 0.5799
Epoch [15/80], Step [100/500] Loss: 0.5295
Epoch [15/80], Step [200/500] Loss: 0.4475
Epoch [15/80], Step [300/500] Loss: 0.3896
Epoch [15/80], Step [400/500] Loss: 0.4869
Epoch [15/80], Step [500/500] Loss: 0.4973
Epoch [20/80], Step [100/500] Loss: 0.3953
Epoch [20/80], Step [200/500] Loss: 0.4542
Epoch [20/80], Step [300/500] Loss: 0.4003
Epoch [20/80], Step [400/500] Loss: 0.3863
Epoch [20/80], Step [500/500] Loss: 0.2813
Epoch [25/80], Step [100/500] Loss: 0.3860
Epoch [25/80], Step [200/500] Loss: 0.4341
Epoch [25/80], Step [300/500] Loss: 0.3384
Epoch [25/80], Step [400/500] Loss: 0.1694
Epoch [25/80], Step [500/500] Loss: 0.2215
Epoch [30/80], Step [100/500] Loss: 0.2096
Epoch [30/80], Step [200/500] Loss: 0.1695
Epoch [30/80], Step [300/500] Loss: 0.2272
Epoch [30/80], Step [400/500] Loss: 0.2907
Epoch [30/80], Step [500/500] Loss: 0.1764
Epoch [35/80], Step [100/500] Loss: 0.2971
Epoch [35/80], Step [200/500] Loss: 0.2568
Epoch [35/80], Step [300/500] Loss: 0.1824
Epoch [35/80], Step [400/500] Loss: 0.1700
Epoch [35/80], Step [500/500] Loss: 0.2449
Epoch [40/80], Step [100/500] Loss: 0.0951
Epoch [40/80], Step [200/500] Loss: 0.2217
Epoch [40/80], Step [300/500] Loss: 0.2020
Epoch [40/80], Step [400/500] Loss: 0.1849
Epoch [40/80], Step [500/500] Loss: 0.1752
Epoch [45/80], Step [100/500] Loss: 0.3183
Epoch [45/80], Step [200/500] Loss: 0.4195
Epoch [45/80], Step [300/500] Loss: 0.2002
Epoch [45/80], Step [400/500] Loss: 0.1956
Epoch [45/80], Step [500/500] Loss: 0.1547
Epoch [50/80], Step [100/500] Loss: 0.2431
Epoch [50/80], Step [200/500] Loss: 0.1655
Epoch [50/80], Step [300/500] Loss: 0.0941
Epoch [50/80], Step [400/500] Loss: 0.2437
Epoch [50/80], Step [500/500] Loss: 0.1340
Epoch [55/80], Step [100/500] Loss: 0.2455
Epoch [55/80], Step [200/500] Loss: 0.1532
Epoch [55/80], Step [300/500] Loss: 0.1303
Epoch [55/80], Step [400/500] Loss: 0.1286
Epoch [55/80], Step [500/500] Loss: 0.2082
Epoch [60/80], Step [100/500] Loss: 0.2705
Epoch [60/80], Step [200/500] Loss: 0.1413
Epoch [60/80], Step [300/500] Loss: 0.1149
Epoch [60/80], Step [400/500] Loss: 0.1146
Epoch [60/80], Step [500/500] Loss: 0.1569
Epoch [65/80], Step [100/500] Loss: 0.1463
Epoch [65/80], Step [200/500] Loss: 0.1799
Epoch [65/80], Step [300/500] Loss: 0.1485
Epoch [65/80], Step [400/500] Loss: 0.1690
Epoch [65/80], Step [500/500] Loss: 0.2135
Epoch [70/80], Step [100/500] Loss: 0.1388
Epoch [70/80], Step [200/500] Loss: 0.1783
Epoch [70/80], Step [300/500] Loss: 0.1284
Epoch [70/80], Step [400/500] Loss: 0.1675
Epoch [70/80], Step [500/500] Loss: 0.2066
Epoch [75/80], Step [100/500] Loss: 0.1681
Epoch [75/80], Step [200/500] Loss: 0.0998
Epoch [75/80], Step [300/500] Loss: 0.1553
Epoch [75/80], Step [400/500] Loss: 0.1153
Epoch [75/80], Step [500/500] Loss: 0.1365
Epoch [80/80], Step [100/500] Loss: 0.1176
Epoch [80/80], Step [200/500] Loss: 0.2006
Epoch [80/80], Step [300/500] Loss: 0.1738
Epoch [80/80], Step [400/500] Loss: 0.1613
Epoch [80/80], Step [500/500] Loss: 0.2003

模型测试和保存

1
2
# 设置为评估模式
model.eval() 
ResNet(
  (conv): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
  (bn): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (relu): ReLU(inplace)
  (layer1): Sequential(
    (0): ResidualBlock(
      (conv1): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (1): ResidualBlock(
      (conv1): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (layer2): Sequential(
    (0): ResidualBlock(
      (conv1): Conv2d(16, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(16, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResidualBlock(
      (conv1): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (layer3): Sequential(
    (0): ResidualBlock(
      (conv1): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResidualBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (avg_pool): AvgPool2d(kernel_size=8, stride=8, padding=0)
  (fc): Linear(in_features=64, out_features=10, bias=True)
)

1
2
3
4
5
6
7
8
9
10
11
12
13
# 屏蔽梯度计算
with torch.no_grad():
    correct = 0
    total = 0
    for images, labels in test_loader:
        images = images.to(device)
        labels = labels.to(device)
        outputs = model(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

    print('Accuracy of the model on the test images: {} %'.format(100 * correct / total))
Accuracy of the model on the test images: 88.24 %

1
2
# 保存模型
torch.save(model.state_dict(), 'resnet.ckpt')

Pytorch模型可视化

导出ONNX模型

1
2
3
4
5
6
import torch.onnx

# 按照输入格式，设计随机输入
dummy_input =torch.randn(1, 3, 32, 32).cuda()   
# 导出模型
torch.onnx.export(model,dummy_input, 'resnet.onnx',verbose=True) 
graph(%0 : Float(1, 3, 32, 32)
      %1 : Float(16, 3, 3, 3)
      %2 : Float(16)
      %3 : Float(16)
      %4 : Float(16)
      %5 : Float(16)
      %6 : Long()
      %7 : Float(16, 16, 3, 3)
      %8 : Float(16)
      %9 : Float(16)
      %10 : Float(16)
      %11 : Float(16)
      %12 : Long()
      %13 : Float(16, 16, 3, 3)
      %14 : Float(16)
      %15 : Float(16)
      %16 : Float(16)
      %17 : Float(16)
      %18 : Long()
      %19 : Float(16, 16, 3, 3)
      %20 : Float(16)
      %21 : Float(16)
      %22 : Float(16)
      %23 : Float(16)
      %24 : Long()
      %25 : Float(16, 16, 3, 3)
      %26 : Float(16)
      %27 : Float(16)
      %28 : Float(16)
      %29 : Float(16)
      %30 : Long()
      %31 : Float(32, 16, 3, 3)
      %32 : Float(32)
      %33 : Float(32)
      %34 : Float(32)
      %35 : Float(32)
      %36 : Long()
      %37 : Float(32, 32, 3, 3)
      %38 : Float(32)
      %39 : Float(32)
      %40 : Float(32)
      %41 : Float(32)
      %42 : Long()
      %43 : Float(32, 16, 3, 3)
      %44 : Float(32)
      %45 : Float(32)
      %46 : Float(32)
      %47 : Float(32)
      %48 : Long()
      %49 : Float(32, 32, 3, 3)
      %50 : Float(32)
      %51 : Float(32)
      %52 : Float(32)
      %53 : Float(32)
      %54 : Long()
      %55 : Float(32, 32, 3, 3)
      %56 : Float(32)
      %57 : Float(32)
      %58 : Float(32)
      %59 : Float(32)
      %60 : Long()
      %61 : Float(64, 32, 3, 3)
      %62 : Float(64)
      %63 : Float(64)
      %64 : Float(64)
      %65 : Float(64)
      %66 : Long()
      %67 : Float(64, 64, 3, 3)
      %68 : Float(64)
      %69 : Float(64)
      %70 : Float(64)
      %71 : Float(64)
      %72 : Long()
      %73 : Float(64, 32, 3, 3)
      %74 : Float(64)
      %75 : Float(64)
      %76 : Float(64)
      %77 : Float(64)
      %78 : Long()
      %79 : Float(64, 64, 3, 3)
      %80 : Float(64)
      %81 : Float(64)
      %82 : Float(64)
      %83 : Float(64)
      %84 : Long()
      %85 : Float(64, 64, 3, 3)
      %86 : Float(64)
      %87 : Float(64)
      %88 : Float(64)
      %89 : Float(64)
      %90 : Long()
      %91 : Float(10, 64)
      %92 : Float(10)) {
  %93 : Float(1, 16, 32, 32) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%0, %1), scope: ResNet/Conv2d[conv]
  %94 : Float(1, 16, 32, 32) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%93, %2, %3, %4, %5), scope: ResNet/BatchNorm2d[bn]
  %95 : Float(1, 16, 32, 32) = onnx::Relu(%94), scope: ResNet/ReLU[relu]
  %96 : Float(1, 16, 32, 32) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%95, %7), scope: ResNet/Sequential[layer1]/ResidualBlock[0]/Conv2d[conv1]
  %97 : Float(1, 16, 32, 32) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%96, %8, %9, %10, %11), scope: ResNet/Sequential[layer1]/ResidualBlock[0]/BatchNorm2d[bn1]
  %98 : Float(1, 16, 32, 32) = onnx::Relu(%97), scope: ResNet/Sequential[layer1]/ResidualBlock[0]/ReLU[relu]
  %99 : Float(1, 16, 32, 32) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%98, %13), scope: ResNet/Sequential[layer1]/ResidualBlock[0]/Conv2d[conv2]
  %100 : Float(1, 16, 32, 32) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%99, %14, %15, %16, %17), scope: ResNet/Sequential[layer1]/ResidualBlock[0]/BatchNorm2d[bn2]
  %101 : Float(1, 16, 32, 32) = onnx::Add(%100, %95), scope: ResNet/Sequential[layer1]/ResidualBlock[0]
  %102 : Float(1, 16, 32, 32) = onnx::Relu(%101), scope: ResNet/Sequential[layer1]/ResidualBlock[0]/ReLU[relu]
  %103 : Float(1, 16, 32, 32) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%102, %19), scope: ResNet/Sequential[layer1]/ResidualBlock[1]/Conv2d[conv1]
  %104 : Float(1, 16, 32, 32) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%103, %20, %21, %22, %23), scope: ResNet/Sequential[layer1]/ResidualBlock[1]/BatchNorm2d[bn1]
  %105 : Float(1, 16, 32, 32) = onnx::Relu(%104), scope: ResNet/Sequential[layer1]/ResidualBlock[1]/ReLU[relu]
  %106 : Float(1, 16, 32, 32) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%105, %25), scope: ResNet/Sequential[layer1]/ResidualBlock[1]/Conv2d[conv2]
  %107 : Float(1, 16, 32, 32) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%106, %26, %27, %28, %29), scope: ResNet/Sequential[layer1]/ResidualBlock[1]/BatchNorm2d[bn2]
  %108 : Float(1, 16, 32, 32) = onnx::Add(%107, %102), scope: ResNet/Sequential[layer1]/ResidualBlock[1]
  %109 : Float(1, 16, 32, 32) = onnx::Relu(%108), scope: ResNet/Sequential[layer1]/ResidualBlock[1]/ReLU[relu]
  %110 : Float(1, 32, 16, 16) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%109, %31), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/Conv2d[conv1]
  %111 : Float(1, 32, 16, 16) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%110, %32, %33, %34, %35), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/BatchNorm2d[bn1]
  %112 : Float(1, 32, 16, 16) = onnx::Relu(%111), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/ReLU[relu]
  %113 : Float(1, 32, 16, 16) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%112, %37), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/Conv2d[conv2]
  %114 : Float(1, 32, 16, 16) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%113, %38, %39, %40, %41), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/BatchNorm2d[bn2]
  %115 : Float(1, 32, 16, 16) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%109, %43), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/Sequential[downsample]/Conv2d[0]
  %116 : Float(1, 32, 16, 16) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%115, %44, %45, %46, %47), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/Sequential[downsample]/BatchNorm2d[1]
  %117 : Float(1, 32, 16, 16) = onnx::Add(%114, %116), scope: ResNet/Sequential[layer2]/ResidualBlock[0]
  %118 : Float(1, 32, 16, 16) = onnx::Relu(%117), scope: ResNet/Sequential[layer2]/ResidualBlock[0]/ReLU[relu]
  %119 : Float(1, 32, 16, 16) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%118, %49), scope: ResNet/Sequential[layer2]/ResidualBlock[1]/Conv2d[conv1]
  %120 : Float(1, 32, 16, 16) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%119, %50, %51, %52, %53), scope: ResNet/Sequential[layer2]/ResidualBlock[1]/BatchNorm2d[bn1]
  %121 : Float(1, 32, 16, 16) = onnx::Relu(%120), scope: ResNet/Sequential[layer2]/ResidualBlock[1]/ReLU[relu]
  %122 : Float(1, 32, 16, 16) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%121, %55), scope: ResNet/Sequential[layer2]/ResidualBlock[1]/Conv2d[conv2]
  %123 : Float(1, 32, 16, 16) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%122, %56, %57, %58, %59), scope: ResNet/Sequential[layer2]/ResidualBlock[1]/BatchNorm2d[bn2]
  %124 : Float(1, 32, 16, 16) = onnx::Add(%123, %118), scope: ResNet/Sequential[layer2]/ResidualBlock[1]
  %125 : Float(1, 32, 16, 16) = onnx::Relu(%124), scope: ResNet/Sequential[layer2]/ResidualBlock[1]/ReLU[relu]
  %126 : Float(1, 64, 8, 8) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%125, %61), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/Conv2d[conv1]
  %127 : Float(1, 64, 8, 8) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%126, %62, %63, %64, %65), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/BatchNorm2d[bn1]
  %128 : Float(1, 64, 8, 8) = onnx::Relu(%127), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/ReLU[relu]
  %129 : Float(1, 64, 8, 8) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%128, %67), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/Conv2d[conv2]
  %130 : Float(1, 64, 8, 8) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%129, %68, %69, %70, %71), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/BatchNorm2d[bn2]
  %131 : Float(1, 64, 8, 8) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%125, %73), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/Sequential[downsample]/Conv2d[0]
  %132 : Float(1, 64, 8, 8) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%131, %74, %75, %76, %77), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/Sequential[downsample]/BatchNorm2d[1]
  %133 : Float(1, 64, 8, 8) = onnx::Add(%130, %132), scope: ResNet/Sequential[layer3]/ResidualBlock[0]
  %134 : Float(1, 64, 8, 8) = onnx::Relu(%133), scope: ResNet/Sequential[layer3]/ResidualBlock[0]/ReLU[relu]
  %135 : Float(1, 64, 8, 8) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%134, %79), scope: ResNet/Sequential[layer3]/ResidualBlock[1]/Conv2d[conv1]
  %136 : Float(1, 64, 8, 8) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%135, %80, %81, %82, %83), scope: ResNet/Sequential[layer3]/ResidualBlock[1]/BatchNorm2d[bn1]
  %137 : Float(1, 64, 8, 8) = onnx::Relu(%136), scope: ResNet/Sequential[layer3]/ResidualBlock[1]/ReLU[relu]
  %138 : Float(1, 64, 8, 8) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%137, %85), scope: ResNet/Sequential[layer3]/ResidualBlock[1]/Conv2d[conv2]
  %139 : Float(1, 64, 8, 8) = onnx::BatchNormalization[epsilon=1e-05, is_test=1, momentum=1](%138, %86, %87, %88, %89), scope: ResNet/Sequential[layer3]/ResidualBlock[1]/BatchNorm2d[bn2]
  %140 : Float(1, 64, 8, 8) = onnx::Add(%139, %134), scope: ResNet/Sequential[layer3]/ResidualBlock[1]
  %141 : Float(1, 64, 8, 8) = onnx::Relu(%140), scope: ResNet/Sequential[layer3]/ResidualBlock[1]/ReLU[relu]
  %142 : Dynamic = onnx::Pad[mode=constant, pads=[0, 0, 0, 0, 0, 0, 0, 0], value=0](%141), scope: ResNet/AvgPool2d[avg_pool]
  %143 : Float(1, 64, 1, 1) = onnx::AveragePool[kernel_shape=[8, 8], pads=[0, 0, 0, 0], strides=[8, 8]](%142), scope: ResNet/AvgPool2d[avg_pool]
  %144 : Dynamic = onnx::Shape(%143), scope: ResNet
  %145 : Dynamic = onnx::Slice[axes=[0], ends=[1], starts=[0]](%144), scope: ResNet
  %146 : Long() = onnx::Squeeze[axes=[0]](%145), scope: ResNet
  %147 : Long() = onnx::Constant[value={-1}](), scope: ResNet
  %148 : Dynamic = onnx::Unsqueeze[axes=[0]](%146), scope: ResNet
  %149 : Dynamic = onnx::Unsqueeze[axes=[0]](%147), scope: ResNet
  %150 : Dynamic = onnx::Concat[axis=0](%148, %149), scope: ResNet
  %151 : Float(1, 64) = onnx::Reshape(%143, %150), scope: ResNet
  %152 : Float(1, 10) = onnx::Gemm[alpha=1, beta=1, broadcast=1, transB=1](%151, %91, %92), scope: ResNet/Linear[fc]
  return (%152);
}

模型可视化工具：NETRON

有几种方式：

  • 安装ONNX客户端
  • ONNX有测试网页可以加载显示模型 ：Netron
  • 安装netron服务，可以通过import netron 和 netron.start('model.onnx')来启动本地查看服务，打开指定端口即可看到。
1
2
3
import netron
#打开服务
netron.start('resnet.onnx')
Stopping http://localhost:8080
Serving 'resnet.onnx' at http://localhost:8080

模型可视化结果

ResNet