python神经网络分析3种Python查看神经网络结构的方法小结pytho

1. 网络结构代码

import torchimport torch.nn as nn 定义Actor-Critic模型class ActorCritic(nn.Module): def __init__(self, state_dim, action_dim): super(ActorCritic, self).__init__() self.actor = nn.Sequential( 全连接层，输入维度为 state_dim，输出维度为 256 nn.Linear(state_dim, 64), nn.ReLU(), nn.Linear(64, action_dim), Softmax 函数，将输出转换为概率分布，dim=-1 表示在最终一个维度上应用 Softmax nn.Softmax(dim=-1) ) self.critic = nn.Sequential( nn.Linear(state_dim, 64), nn.ReLU(), nn.Linear(64, 1) ) def forward(self, state): policy = self.actor(state) value = self.critic(state) return policy, value 参数设置state_dim = 1action_dim = 2model = ActorCritic(state_dim, action_dim)

2. 查看结构

2.1 直接打印模型

print(model)

输出：

ActorCritic(
(actor): Sequential(
(0): Linear(in_features=1, out_features=64, bias=True)
(1): ReLU()
(2): Linear(in_features=64, out_features=2, bias=True)
(3): Softmax(dim=-1)
)
(critic): Sequential(
(0): Linear(in_features=1, out_features=64, bias=True)
(1): ReLU()
(2): Linear(in_features=64, out_features=1, bias=True)
)
)

2.2 可视化网络结构（需要安装 torchviz 包）

安装 torchsummary 包：

$ pip install torchsummary

python 代码：

from torchviz import make_dot 创建一个虚拟输入x = torch.randn(1, state_dim) 生成计算图dot = make_dot(model(x), params=dict(model.named_parameters()))dot.render(“actor_critic_model”, format=”png”) 保存为PNG图片

输出 actor_critic_model

digraph
graph [size="12,12"]
node [align=left fontname=monospace fontsize=10 height=0.2 ranksep=0.1 shape=box style=filled]
140281544075344 [label="
(1, 2)" fillcolor=darkolivegreen1]
140281544213744 [label=SoftmaxBackward0]
140281544213840 -> 140281544213744
140281544213840 [label=AddmmBackward0]
140281544213600 -> 140281544213840
140285722327344 [label="actor.2.bias
(2)" fillcolor=lightblue]
140285722327344 -> 140281544213600
140281544213600 [label=AccumulateGrad]
140281544214032 -> 140281544213840
140281544214032 [label=ReluBackward0]
140281544213984 -> 140281544214032
140281544213984 [label=AddmmBackward0]
140281544214176 -> 140281544213984
140285722327024 [label="actor.0.bias
(64)" fillcolor=lightblue]
140285722327024 -> 140281544214176
140281544214176 [label=AccumulateGrad]
140281544214224 -> 140281544213984
140281544214224 [label=TBackward0]
140281543934832 -> 140281544214224
140285722327264 [label="actor.0.weight
(64, 1)" fillcolor=lightblue]
140285722327264 -> 140281543934832
140281543934832 [label=AccumulateGrad]
140281544213648 -> 140281544213840
140281544213648 [label=TBackward0]
140281544214080 -> 140281544213648
140285722327184 [label="actor.2.weight
(2, 64)" fillcolor=lightblue]
140285722327184 -> 140281544214080
140281544214080 [label=AccumulateGrad]
140281544213744 -> 140281544075344
140285722328704 [label="
(1, 1)" fillcolor=darkolivegreen1]
140281544213888 [label=AddmmBackward0]
140281544214368 -> 140281544213888
140285722328064 [label="critic.2.bias
(1)" fillcolor=lightblue]
140285722328064 -> 140281544214368
140281544214368 [label=AccumulateGrad]
140281544214128 -> 140281544213888
140281544214128 [label=ReluBackward0]
140281544214464 -> 140281544214128
140281544214464 [label=AddmmBackward0]
140281544214512 -> 140281544214464
140285722327424 [label="critic.0.bias
(64)" fillcolor=lightblue]
140285722327424 -> 140281544214512
140281544214512 [label=AccumulateGrad]
140281544214560 -> 140281544214464
140281544214560 [label=TBackward0]
140281544214704 -> 140281544214560
140285722327504 [label="critic.0.weight
(64, 1)" fillcolor=lightblue]
140285722327504 -> 140281544214704
140281544214704 [label=AccumulateGrad]
140281544213696 -> 140281544213888
140281544213696 [label=TBackward0]
140281544214272 -> 140281544213696
140285722327584 [label="critic.2.weight
(1, 64)" fillcolor=lightblue]
140285722327584 -> 140281544214272
140281544214272 [label=AccumulateGrad]
140281544213888 -> 140285722328704
}

输出模型图片：

2.3 使用 summary 技巧（需要安装 torchsummary 包）

安装 torchsummary 包：

pip install torchsummary

代码：

from torchsummary import summarydevice = torch.device(“cuda:0” if torch.cuda.is_available() else “cpu”)print(device)model = model.to(device)summary(model, input_size=(state_dim,))查看模型参数print(“查看模型参数：”)for name, param in model.named_parameters(): print(f”Layer: name} | Size: param.size()} | Values: param[:2]}…”)

输出：

cuda:0
&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;-
Layer (type) Output Shape Param
================================================================
Linear-1 [-1, 64] 128
ReLU-2 [-1, 64] 0
Linear-3 [-1, 2] 130
Softmax-4 [-1, 2] 0
Linear-5 [-1, 64] 128
ReLU-6 [-1, 64] 0
Linear-7 [-1, 1] 65
================================================================
Total params: 451
Trainable params: 451
Non-trainable params: 0
&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;-
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 0.00
Estimated Total Size (MB): 0.00
&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;&8212;-
查看模型参数：
Layer: actor.0.weight | Size: torch.Size([64, 1]) | Values: tensor([[ 0.7747],
[-0.0440]], device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: actor.0.bias | Size: torch.Size([64]) | Values: tensor([ 0.5995, -0.2155], device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: actor.2.weight | Size: torch.Size([2, 64]) | Values: tensor([[ 0.0373, 0.0851, 0.1000, 0.1060, 0.0387, 0.0479, 0.0127, 0.0696,
0.0388, 0.0033, 0.1173, -0.1195, -0.0830, 0.0186, 0.0063, -0.0863,
-0.0353, 0.0782, -0.0558, 0.0011, -0.0533, 0.1241, 0.0120, -0.0906,
-0.0551, -0.0673, -0.1070, 0.0402, -0.0662, 0.0596, -0.0811, 0.0457,
0.0349, 0.0564, -0.0155, -0.0404, 0.0843, -0.0978, 0.0459, 0.1097,
-0.0858, 0.0736, -0.0067, -0.0756, -0.0363, -0.0525, -0.0426, -0.1087,
-0.0611, 0.0420, -0.1038, 0.0402, 0.0065, -0.1217, -0.0467, 0.0383,
-0.0217, 0.0283, 0.0800, 0.0228, 0.0415, -0.0473, -0.0199, -0.0436],
[-0.1118, -0.0806, -0.0700, -0.0224, 0.0335, -0.0087, 0.0265, -0.1196,
-0.0907, -0.0360, 0.0621, -0.0471, -0.0939, -0.0912, -0.1061, 0.1051,
-0.0592, -0.0757, 0.0758, -0.1082, -0.0317, 0.1208, -0.0279, -0.0693,
0.0920, -0.0318, -0.0476, 0.0236, -0.0761, 0.0591, 0.0862, -0.0712,
0.0156, -0.1073, 0.1133, 0.0039, -0.0191, 0.0605, -0.0686, -0.1202,
0.0962, 0.0581, 0.1145, 0.0741, -0.0993, -0.0987, 0.0939, 0.1006,
0.0773, -0.0756, -0.1096, 0.0156, -0.0599, 0.0857, 0.1005, -0.0618,
0.0474, 0.0066, -0.0531, -0.0479, 0.1136, 0.0356, 0.1169, -0.0023]],
device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: actor.2.bias | Size: torch.Size([2]) | Values: tensor([-0.0039, 0.0937], device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: critic.0.weight | Size: torch.Size([64, 1]) | Values: tensor([[0.5799],
[0.0473]], device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: critic.0.bias | Size: torch.Size([64]) | Values: tensor([ 0.6507, -0.6974], device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: critic.2.weight | Size: torch.Size([1, 64]) | Values: tensor([[ 0.0738, -0.0370, -0.1010, -0.0333, -0.0595, -0.0172, 0.0928, 0.0815,
0.1221, -0.0842, 0.0511, 0.0452, -0.0386, -0.0503, -0.0964, 0.0370,
-0.0341, -0.0693, -0.0845, 0.0424, -0.0491, -0.0439, -0.0443, 0.0203,
0.0960, -0.1178, -0.0836, -0.0144, -0.0576, -0.0851, 0.0461, 0.1160,
0.0120, 0.1180, 0.0255, 0.1047, -0.0398, 0.0786, 0.1143, 0.0806,
0.1125, 0.0267, 0.0534, -0.0318, 0.1125, -0.0727, 0.1169, 0.0120,
-0.0178, -0.0845, 0.0069, 0.0194, 0.1188, 0.0481, 0.1077, -0.0840,
0.1013, 0.0586, -0.0857, -0.0974, -0.0630, 0.0359, -0.0080, -0.0926]],
device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;
Layer: critic.2.bias | Size: torch.Size([1]) | Values: tensor([0.0621], device=&039;cuda:0&039;, grad_fn=<SliceBackward0>)&8230;

到此这篇关于3种Python查看神经网络结构的技巧的文章就介绍到这了,更多相关Python查看神经网络结构内容请搜索风君子博客以前的文章或继续浏览下面的相关文章希望大家以后多多支持风君子博客！

无论兄弟们可能感兴趣的文章:

Python基于numpy灵活定义神经网络结构的技巧
Python调用graphviz绘制结构化图形网络示例
python目标检测yolo1?yolo2?yolo3和SSD网络结构对比

英网客