PyTorch 打印模型结构、输出维度和参数信息(torchsummary)
发布时间:2024-02-11 01:05:00

PyTorch 打印模型结构、输出维度和参数信息(torchsummary)

使用 PyTorch 深度学习搭建模型后,如果想查看模型结构,可以直接使用 print(model) 函数打印。但该输出结果不是特别直观,查阅发现有个能输出类似 keras 风格 model.summary() 的模型可视化工具。这里记录一下方便以后查阅。

安装 torchsummary

pip install torchsummary

输出网络信息

summary函数介绍

model:网络模型
input_size:网络输入图片的shape,这里不用加batch_size进去
batch_size:batch_size参数,默认是-1
device:在GPU还是CPU上运行,默认是cuda在GPU上运行,如果想在CPU上执行将参数改为CPU即可

import torch
import torch.nn as nn
from torchsummary import summary



class Shallow_ConvNet(nn.Module):
    def __init__(self, in_channel, conv_channel_temp, kernel_size_temp, conv_channel_spat, kernel_size_spat,
                              pooling_size, pool_stride_size, dropoutRate, n_classes, class_kernel_size) :
        super(Shallow_ConvNet, self).__init__()

        self.temp_conv = nn.Conv2d(in_channels=in_channel,
                                                                    out_channels=conv_channel_temp,
                                                                    kernel_size=(1, kernel_size_temp),
                                                                    stride=1,
                                                                    bias=False)

        self.spat_conv = nn.Conv2d(in_channels=conv_channel_temp,
                                                                  out_channels=conv_channel_spat,
                                                                  kernel_size=(kernel_size_spat, 1),
                                                                  stride=1,
                                                                  bias=False)

        self.bn = nn.BatchNorm2d(num_features=conv_channel_spat)

        # slef.act_conv = x*x

        self.pooling = nn.AvgPool2d(kernel_size=(1, pooling_size),
                                                                   stride=(1, pool_stride_size))

        # slef.act_pool = log(max(x, eps))

        self.dropout = nn.Dropout(p=dropoutRate)

        self.class_conv = nn.Conv2d(in_channels=conv_channel_spat,
                                                                    out_channels=n_classes,
                                                                    kernel_size=(1, class_kernel_size),
                                                                    bias=False)

        self.softmax = nn.Softmax(dim=1)

    def safe_log(self, x):
        """ Prevents :math:`log(0)` by using :math:`log(max(x, eps))`."""
        return torch.log(torch.clamp(x, min=1e-6))
    
    def forward(self, x):
        # input shape (batch_size, C, T)
        if len(x.shape) is not 4:
            x = torch.unsqueeze(x, 1)
        # input shape (batch_size, 1, C, T)
        x = self.temp_conv(x)
        x = self.spat_conv(x)
        x = self.bn(x)
        x = x*x # conv_activate
        x = self.pooling(x)
        x = self.safe_log(x) # pool_activate
        x = self.dropout(x)
        x = self.class_conv(x)
        x= self.softmax(x)
        out = torch.squeeze(x)

        return out


###============================ Initialization parameters ============================###
channels = 44
samples = 534

in_channel = 1
conv_channel_temp = 40
kernel_size_temp = 25
conv_channel_spat = 40
kernel_size_spat = channels
pooling_size = 75
pool_stride_size = 15
dropoutRate = 0.3
n_classes = 4
class_kernel_size = 30

def main():
    input = torch.randn(32, 1, channels, samples)
    model = Shallow_ConvNet(in_channel, conv_channel_temp, kernel_size_temp, conv_channel_spat, kernel_size_spat,
                                                            pooling_size, pool_stride_size, dropoutRate, n_classes, class_kernel_size)
    out = model(input)
    print('===============================================================')
    print('out', out.shape)
    print('model', model)
    summary(model=model, input_size=(1,channels,samples), batch_size=32, device="cpu")

if __name__ == "__main__":
    main()

输出:

out torch.Size([32, 4])
model Shallow_ConvNet(
  (temp_conv): Conv2d(1, 40, kernel_size=(1, 25), stride=(1, 1), bias=False)
  (spat_conv): Conv2d(40, 40, kernel_size=(44, 1), stride=(1, 1), bias=False)
  (bn): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (pooling): AvgPool2d(kernel_size=(1, 75), stride=(1, 15), padding=0)
  (dropout): Dropout(p=0.3, inplace=False)
  (class_conv): Conv2d(40, 4, kernel_size=(1, 30), stride=(1, 1), bias=False)
  (softmax): Softmax(dim=1)
)
----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1          [32, 40, 44, 510]           1,000
            Conv2d-2           [32, 40, 1, 510]          70,400
       BatchNorm2d-3           [32, 40, 1, 510]              80
         AvgPool2d-4            [32, 40, 1, 30]               0
           Dropout-5            [32, 40, 1, 30]               0
            Conv2d-6              [32, 4, 1, 1]           4,800
           Softmax-7              [32, 4, 1, 1]               0
================================================================
Total params: 76,280
Trainable params: 76,280
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 2.87
Forward/backward pass size (MB): 229.69
Params size (MB): 0.29
Estimated Total Size (MB): 232.85
----------------------------------------------------------------

AttributeError: ‘tuple’ object has no attribute ‘size’

旧的summary加入LSTM之类的会报错,需要用新的summarry

pip install torchinfo
from torchinfo import summary

def main():
    input = torch.randn(32, window_size, channels, samples)
    model = Cascade_Conv_LSTM(in_channel, out_channel_conv1, out_channel_conv2, out_channel_conv3, kernel_conv123, stride_conv123, padding_conv123,
                                                                    fc1_in, fc1_out, dropoutRate1, lstm1_in, lstm1_hidden, lstm1_layer, lstm2_in, lstm2_hidden, lstm2_layer, fc2_in, fc2_out, dropoutRate2,
                                                                    fc3_in, n_classes)
    # model = model.to('cuda:1')
    # input = torch.from_numpy(input).to('cuda:1').to(torch.float32).requires_grad_()
    out = model(input)
    print('===============================================================')
    print('out', out.shape)
    print('model', model)
    summary(model=model, input_size=(32,10,channels,samples), device="cpu")

if __name__ == "__main__":
    main()
==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
Cascade_Conv_LSTM                        [32, 4]                   --
├─Sequential: 1-1                        [320, 32, 10, 11]         --
│    └─Conv2d: 2-1                       [320, 32, 10, 11]         288
│    └─ELU: 2-2                          [320, 32, 10, 11]         --
├─Sequential: 1-2                        [320, 64, 10, 11]         --
│    └─Conv2d: 2-3                       [320, 64, 10, 11]         18,432
│    └─ELU: 2-4                          [320, 64, 10, 11]         --
├─Sequential: 1-3                        [320, 128, 10, 11]        --
│    └─Conv2d: 2-5                       [320, 128, 10, 11]        73,728
│    └─ELU: 2-6                          [320, 128, 10, 11]        --
├─Sequential: 1-4                        [320, 1024]               --
│    └─Linear: 2-7                       [320, 1024]               14,418,944
│    └─ELU: 2-8                          [320, 1024]               --
├─Dropout: 1-5                           [320, 1024]               --
├─LSTM: 1-6                              [32, 10, 1024]            8,396,800
├─LSTM: 1-7                              [32, 10, 1024]            8,396,800
├─Sequential: 1-8                        [32, 1024]                --
│    └─Linear: 2-9                       [32, 1024]                1,049,600
│    └─ELU: 2-10                         [32, 1024]                --
├─Dropout: 1-9                           [32, 1024]                --
├─Linear: 1-10                           [32, 4]                   4,100
├─Softmax: 1-11                          [32, 4]                   --
==========================================================================================
Total params: 32,358,692
Trainable params: 32,358,692
Non-trainable params: 0
Total mult-adds (G): 13.28
==========================================================================================
Input size (MB): 0.14
Forward/backward pass size (MB): 71.21
Params size (MB): 129.43
Estimated Total Size (MB): 200.78
==========================================================================================