目录
FNN 前馈神经网络
网络设计
代码实现
因为mnist中的所有图片都是28*28*1的图片，1表示channel，就是灰色图片
import torch # Import PyTorch
from torch import nn # Import the neural network module from PyTorch
from torchinfo import summary
# Define the neural network class, inheriting from nn.Module
class PreNetwork(nn.Module):
def __init__(self):
super().__init__() # Call the initializer of the parent class nn.Module
self.layer1 = nn.Linear(784, 256) # Define the first linear layer (input size 784, output size 256)
self.layer2 = nn.Linear(256, 10) # Define the second linear layer (input size 256, output size 10)
def forward(self, x):
x = x.view(-1, 28*28) # Flatten the input tensor to a 1D tensor of size 28*28
x = self.layer1(x) # Pass the input through the first linear layer
x = torch.relu(x) # Apply the ReLU activation function
return self.layer2(x) # Pass the result through the second linear layer and return it
验证模型是否正常
net = PreNetwork()
summary(net, input_size=(1,1, 28, 28))测试结果
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
PreNetwork [1, 10] --
├─Linear: 1-1 [1, 256] 200,960
├─Linear: 1-2 [1, 10] 2,570
==========================================================================================
Total params: 203,530
Trainable params: 203,530
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 0.20
==========================================================================================
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 0.81
Estimated Total Size (MB): 0.82
==========================================================================================LeNet卷积神经网络
网络设计
代码设计
从上面的设计图上可以看到，输入的图片是32*32*1，但是mnist的图片是28*28*1，那么如果按照上面的设计进行写代码时，一定会出现问题。
import torch.nn as nn
import torch.nn.functional as func
from torchinfo import summary
class LeNet5(nn.Module):
def __init__(self):
super(LeNet5, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(1, 6, kernel_size=5),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
nn.Conv2d(6, 16, kernel_size=5),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
)
self.classifier = nn.Sequential(
nn.Linear(16*5*5, 120),
nn.Linear(120, 84),
nn.Linear(84, 10),
)
def forward(self, x):
x = self.features(x)
x = x.view(x.size(0), -1)
x = self.classifier(x)
return x
测试
net = LeNet5()
summary(net, input_size=(1,1, 28, 28))测试结果
从测试结果可以看到出现错误
从上面的设计图上可以到，数据经过2个卷积层和2个池化层，然后3个全连接层。
排查问题的方法
1.保证输入的图片channel 一致，例如mnist是 1channel，那么确认一下，输入层的channel是1
nn.Conv2d(1, 6, kernel_size=5)
2.中间的卷积层和池化层不会出问题，这时，就需要看看池化层和全连接层的地方是否有问题。
如果排查，可以先把 x = self.classifier(x) 注释一下
import torch.nn as nn
import torch.nn.functional as func
from torchinfo import summary
class LeNet5(nn.Module):
def __init__(self):
super(LeNet5, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(1, 6, kernel_size=5),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
nn.Conv2d(6, 16, kernel_size=5),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
)
self.classifier = nn.Sequential(
nn.Linear(16*5*5, 120),
nn.Linear(120, 84),
nn.Linear(84, 10),
)
def forward(self, x):
x = self.features(x)
x = x.view(x.size(0), -1)
#x = self.classifier(x)
return x
再次运行测试
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
LeNet5 [1, 256] 41,854
├─Sequential: 1-1 [1, 16, 4, 4] --
│ └─Conv2d: 2-1 [1, 6, 24, 24] 156
│ └─ReLU: 2-2 [1, 6, 24, 24] --
│ └─MaxPool2d: 2-3 [1, 6, 12, 12] --
│ └─Conv2d: 2-4 [1, 16, 8, 8] 2,416
│ └─ReLU: 2-5 [1, 16, 8, 8] --
│ └─MaxPool2d: 2-6 [1, 16, 4, 4] --
==========================================================================================
Total params: 44,426
Trainable params: 44,426
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 0.24
==========================================================================================
Input size (MB): 0.00
Forward/backward pass size (MB): 0.04
Params size (MB): 0.01
Estimated Total Size (MB): 0.05
==========================================================================================可以看到，经过两次卷积和两次池化后，最后的数据为：16*4*4，但是我们上面的代码写的16*5*5,所以这里更改为16*4*4.更改后再进行测试
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
LeNet5 [1, 10] --
├─Sequential: 1-1 [1, 16, 4, 4] --
│ └─Conv2d: 2-1 [1, 6, 24, 24] 156
│ └─ReLU: 2-2 [1, 6, 24, 24] --
│ └─MaxPool2d: 2-3 [1, 6, 12, 12] --
│ └─Conv2d: 2-4 [1, 16, 8, 8] 2,416
│ └─ReLU: 2-5 [1, 16, 8, 8] --
│ └─MaxPool2d: 2-6 [1, 16, 4, 4] --
├─Sequential: 1-2 [1, 10] --
│ └─Linear: 2-7 [1, 120] 30,840
│ └─Linear: 2-8 [1, 84] 10,164
│ └─Linear: 2-9 [1, 10] 850
==========================================================================================
Total params: 44,426
Trainable params: 44,426
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 0.29
==========================================================================================
Input size (MB): 0.00
Forward/backward pass size (MB): 0.04
Params size (MB): 0.18
Estimated Total Size (MB): 0.22
==========================================================================================AlexNet神经网络
网络设计
代码设计
import torch.nn as nn
'''
modified to fit dataset size
'''
NUM_CLASSES = 10
class AlexNet(nn.Module):
def __init__(self, num_classes=NUM_CLASSES):
super(AlexNet, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(1, 64, kernel_size=3, stride=2, padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
nn.Conv2d(64, 192, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
nn.Conv2d(192, 384, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(384, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2),
)
self.classifier = nn.Sequential(
nn.Dropout(),
nn.Linear(256 * 1 * 1, 4096),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Linear(4096, 4096),
nn.ReLU(inplace=True),
nn.Linear(4096, num_classes),
)
def forward(self, x):
x = self.features(x)
x = x.view(x.size(0), -1)
x = self.classifier(x)
return x
测试模型
net = AlexNet()
summary(net, input_size=(20,1, 28, 28))测试结果
从下面的数据中很容易看到，模型中，有5个卷积层和3个池化层和3个全连接层，其中还加入和dropout的功能
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
AlexNet [20, 10] --
├─Sequential: 1-1 [20, 256, 1, 1] --
│ └─Conv2d: 2-1 [20, 64, 14, 14] 640
│ └─ReLU: 2-2 [20, 64, 14, 14] --
│ └─MaxPool2d: 2-3 [20, 64, 7, 7] --
│ └─Conv2d: 2-4 [20, 192, 7, 7] 110,784
│ └─ReLU: 2-5 [20, 192, 7, 7] --
│ └─MaxPool2d: 2-6 [20, 192, 3, 3] --
│ └─Conv2d: 2-7 [20, 384, 3, 3] 663,936
│ └─ReLU: 2-8 [20, 384, 3, 3] --
│ └─Conv2d: 2-9 [20, 256, 3, 3] 884,992
│ └─ReLU: 2-10 [20, 256, 3, 3] --
│ └─Conv2d: 2-11 [20, 256, 3, 3] 590,080
│ └─ReLU: 2-12 [20, 256, 3, 3] --
│ └─MaxPool2d: 2-13 [20, 256, 1, 1] --
├─Sequential: 1-2 [20, 10] --
│ └─Dropout: 2-14 [20, 256] --
│ └─Linear: 2-15 [20, 4096] 1,052,672
│ └─ReLU: 2-16 [20, 4096] --
│ └─Dropout: 2-17 [20, 4096] --
│ └─Linear: 2-18 [20, 4096] 16,781,312
│ └─ReLU: 2-19 [20, 4096] --
...
Input size (MB): 0.06
Forward/backward pass size (MB): 6.11
Params size (MB): 80.50
Estimated Total Size (MB): 86.68
==========================================================================================
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
