Pytorch的网络结构可视化:Netron

最近刚刚发现一个非常好用的显示模型神器Netron

https://github.com/lutzroeder/Netron

https://www.electronjs.org/apps/netron

借助这个工具可以像windows的软件一样导入已经训练好的模型加权重即可一键生成

以下是我的一个模型使用该工具可视化结果，只不过目前该工具对于onnx支持非常好，但是pytorch权重转变为onnx是非常方便的,只需要torch.onnx.export()命令即可导出onnx权重

以下代码将resnet18-5c106cde.pth转化为resnet18.onnx

# -- coding: utf-8 --
import io
import torch
import torch.onnx
import torchvisiondevice = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")def pth2onnx():model = torchvision.models.resnet18(pretrained=False)# pth_file = 'resnet18-5c106cde.pth'# model.load_state_dict(torch.load(pth_file, map_location=device))model.to(device)# data type n*c*h*wdummy_input = torch.randn(1, 3, 256, 256)dummy_input = dummy_input.to(device)input_names = ['actual_input']output_names = ['output']torch.onnx.export(model, dummy_input, 'resnet18.onnx', verbose=True, input_names=input_names,output_names=output_names)if __name__ == '__main__':pth2onnx()

resnet18网络结构图

我目前看了下visdom实现pytorch的网络结构查找还是很困难，在stackflow上有很多人使用自己编写的基于matplotlib来实现网络结构可视化适用性也不是很好，后来查找到使用基于tensorboard所开发的tensorboardX可以很方便的实现pytorch网络结构的可视化，因此决定采用这种方式。

Pytorch的网络结构可视化:TensorBoardX

1.TensorBoardX简介

tensorboardX的项目路径：https://github.com/lanpa/tensorboardX

tensorboardX是基于tensorboard的思想用来写tensorboard events的工具，可以实现对传统的tensorboard中 scalar，image，figure，histogram，audio，text，graph，onnx_graph等事件进行编写。

tensorboardX同时具有论坛供大家提出问题解决问题，论坛地址：https://github.com/lanpa/tensorboardX/wiki

2. tensorboardX的使用

tensorboardX的安装以及依赖如下所示：

pip install tensorboardpip install tensorflowpip install tensorboardX

tensorboardX的路径下带的有一个规范的demo，可以供大家参考。我这里公布一个我自己测试过的代码，代码来源于：https://blog.csdn.net/sunqiande88/article/details/80155925

# -- coding: utf-8 --
import torch
import torch.nn as nn
from tensorboardX import SummaryWriterclass LeNet(nn.Module):def __init__(self):super(LeNet, self).__init__()self.conv1 = nn.Sequential(  # input size=(1*28*28)nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5, stride=1, padding=2),nn.ReLU(),  # (6*28*28)nn.MaxPool2d(kernel_size=2, stride=2),  # output_size(6*14*14))self.conv2 = nn.Sequential(nn.Conv2d(in_channels=6, out_channels=16, kernel_size=5),nn.ReLU(),  # (16*10*10)nn.MaxPool2d(kernel_size=2, stride=2)  # output_size=(16*5*5))self.fc1 = nn.Sequential(nn.Linear(in_features=16 * 5 * 5, out_features=120),nn.ReLU())self.fc2 = nn.Sequential(nn.Linear(in_features=120, out_features=84),nn.ReLU())self.fc3 = nn.Linear(84, 10)def forward(self, x):x = self.conv1(x)x = self.conv2(x)# Linear的输入和输出都是维度为1的值，所以要把多维度的tensor展平成一维x = x.view(x.size()[0], -1)x = self.fc1(x)x = self.fc2(x)x = self.fc3(x)return xdummy_input = torch.rand(4, 1, 28, 28)  # 假设输入4张1*28*28的图片
model = LeNet()
with SummaryWriter(comment='LeNet') as w:w.add_graph(model, (dummy_input,))

运行该代码后会自动生成一个runs文件夹，并且在文件夹下会有一个对应的event，如下图所示：

此时需要在terminal或者cmd下运行tensorboard --logdir = path
此处千万要注意，如果按照上面的参考文档的方式是会报错的：No graph definition files were found 或者 No definition files were found，总之无法正常显示网络结构图。

此处的path 是event对应的确切，完整的路径

在运行后会出现一个http url，此时需要将该url 拷贝到chrome下即可看到如下所示框图

结构框图如下所示：

此时双击红圈所示的LeNet模块即可看到LeNet的细节信息，如下所示：

至此即可完成使用tensorboardX 对pytorch网络结构的可视化

TensorBoard is not found.

Just started using Tensorflow, but I am not able to use tensorboard command on my cmd, it gives the error command

C:\Users\tushar\PycharmProjects>tensorboard --logdir="NewTF"
'tensorboard' is not recognized as an internal or external command,operable program or batch file.

I had the same problem for tensorflow 1.5.0 and windows10.

Following tensor documentation (“Launching TensorBoard” section), you can try:

python -m tensorboard.main --logdir=[PATH_TO_LOGDIR]

参考资料

https://blog.csdn.net/xiaoxifei/article/details/82735355
https://blog.csdn.net/lei19880402/article/details/103721362