pytorch网络模型的可视化总结

方法一：

自己下载graphviz程序，然后编写make_dot函数，然后进行调用。
下载graphviz程序可以参考：文章一、文章二

参考的demo如下：

import torch
from torch.autograd import Variable
import torch.nn as nn
from graphviz import Digraph


class CNN(nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=1, out_channels=16, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(in_channels=16, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.out = nn.Linear(32 * 7 * 7, 10)


    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size(0), -1)  # (batch, 32*7*7)
        out = self.out(x)
        return out


def make_dot(var, params=None):
    """
    画出 PyTorch 自动梯度图 autograd graph 的 Graphviz 表示.
    蓝色节点表示有梯度计算的变量Variables;
    橙色节点表示用于 torch.autograd.Function 中的 backward 的张量 Tensors.
    Args:
        var: output Variable
        params: dict of (name, Variable) to add names to node that
            require grad (TODO: make optional)
    """
    if params is not None:
        assert all(isinstance(p, Variable) for p in params.values())
        param_map = {id(v): k for k, v in params.items()}

    node_attr = dict(style='filled', shape='box', align='left',
                     fontsize='12', ranksep='0.1', height='0.2')
    dot = Digraph(node_attr=node_attr, graph_attr=dict(size="12,12"))
    seen = set()

    def size_to_str(size):
        return '(' + (', ').join(['%d' % v for v in size]) + ')'

    output_nodes = (var.grad_fn,) if not isinstance(var, tuple) else tuple(v.grad_fn for v in var)

    def add_nodes(var):
        if var not in seen:
            if torch.is_tensor(var):
                # note: this used to show .saved_tensors in pytorch0.2, but stopped
                # working as it was moved to ATen and Variable-Tensor merged
                dot.node(str(id(var)), size_to_str(var.size()), fillcolor='orange')
            elif hasattr(var, 'variable'):
                u = var.variable
                name = param_map[id(u)] if params is not None else ''
                node_name = '%s\n %s' % (name, size_to_str(u.size()))
                dot.node(str(id(var)), node_name, fillcolor='lightblue')
            elif var in output_nodes:
                dot.node(str(id(var)), str(type(var).__name__), fillcolor='darkolivegreen1')
            else:
                dot.node(str(id(var)), str(type(var).__name__))
            seen.add(var)
            if hasattr(var, 'next_functions'):
                for u in var.next_functions:
                    if u[0] is not None:
                        dot.edge(str(id(u[0])), str(id(var)))
                        add_nodes(u[0])
            if hasattr(var, 'saved_tensors'):
                for t in var.saved_tensors:
                    dot.edge(str(id(t)), str(id(var)))
                    add_nodes(t)

    # 多输出场景 multiple outputs
    if isinstance(var, tuple):
        for v in var:
            add_nodes(v.grad_fn)
    else:
        add_nodes(var.grad_fn)
    return dot

if __name__ == '__main__':
    net = CNN()
    x = torch.randn(1, 1, 28, 28)
    y = net(x)
    g = make_dot(y)
    g.view()
    params = list(net.parameters())
    k = 0
    for i in params:
        l = 1
        print("该层的结构：" + str(list(i.size())))
        for j in i.size():
            l *= j
            print("该层参数和：" + str(l))
            k = k + l
            print("总参数数量和：" + str(k))

方法二：

直接使用torchviz，使用前先用pip进行安装即可。
其实这种方法和方法一一样，只不过是方法一把torchviz.make_dot单独copy出来了。

参考的demo如下：

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchviz import make_dot

class CNN(nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=1, out_channels=16, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(in_channels=16, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.out = nn.Linear(32 * 7 * 7, 10)


    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size(0), -1)  # (batch, 32*7*7)
        out = self.out(x)
        return out

net = CNN()
print(net)
x = torch.zeros(1, 1, 28, 28, dtype=torch.float, requires_grad=True)
net_out = net(x)
net_struct = make_dot(net_out)  # plot graph of variable, not of a nn.Module
net_struct.view()
# net_struct.render("net_struct", view=True)

方法三：

Netron开源地址： https://github.com/lutzroeder/Netron
Netron使用很简单，但是功能却很强大。作者提供了各个平台的安装包，安装之后打开，把保存的模型文件拖入就可以了。

import torch
from torch import nn
from torchviz import make_dot, make_dot_from_trace
 
model = nn.Sequential()
model.add_module('W0', nn.Linear(8, 16))
model.add_module('tanh', nn.Tanh())
model.add_module('W1', nn.Linear(16, 1))
 
torch.save(model, 'model.pth')  # 保存模型

之后用Netron打开保存的“model.pth”即可。
如果你懒得安装，还可以使用作者提供的在线Netron查看器，地址：https://lutzroeder.github.io/netron/