PyCharm入门学习笔记（三）神经网络的基本骨架torch.nn

一、神经网络（Neural network）架构torch.nn

1.1 神经网络架构

1.2 卷积神经网络

1.2.1 介绍

1. 卷积神经网络（CNN）应用于计算机视觉领域，特征提取方法，任务：图像的分类和检索。
2. GPU图像处理单元：GPU做卷积比CPU快得多
3. 卷积网络与传统神经网络区别：数据二维->三维
4. 整体架构：
   

1.2.2 输入层

1.2.3 卷积层

1. 卷积的任务：图像分割成很多部分，对于不同区域提取出不同特征，经过卷积核计算，计算出每个区域的特征值，将5×5×5-> 3×3×3，卷积可做多次
   

2. 图像颜色通道：3个颜色通道R G B做加法
   

3. 输入与不同的卷积核Filter Wi做内积得到的结果不同，一般做完卷积会加上一个偏置参数bias（下图对应：in_channel=3，out_channel=2，卷积核个数/大小为2/3×3，第一/二个bias为1/0）
   

4. 卷积层一些参数：
   1）滑动窗口步长：不同步长得到的输出不同，一般来说步长越小精度越高，但效率会变慢，一般选择1
   2）卷积核尺寸：Filter Wi的尺寸，一般选择3×3
   3）边缘填充：往上数第二图的输入的图，边界的点被计算的次数少，显然不公平。于是在边界点外侧再填充一些点（灰色部分0）
   4）卷积核个数（n）：Filter Wi i=0,1…,n-1。要想得到多个特征图，需要设置不同的卷积核。
   例：11×11 fileters at stride 4 pade 0表示卷积核尺寸11×11，步长2，填充0
   
   
   
   

1.2.4 池化层：压缩 筛选

1.2.5 全连接层：卷积层和池化层只做特征提取，全连接层将一个图拉成特征向量

例如：
32×32×10->[10240,5]
带参数计算的才算做一层，卷积（CONV计算） 激活（RELU无计算） 池化（POOL无计算） 全连接（FC计算），所以以下7层神经网络。

1.2.6 经典的架构：

1. 12年Alexnet：8层
   

2. 14年Vgg：16/19层，且卷积核3×3，经过maxpolling损失的信息在下一步弥补起来
   

3. 15年Resnet：按理说层数越高效果越好，但是实际并不是这样。于是，提出“同等映射”，在堆叠过程中得到F(x)如果表现不好将F(x)变为0，再在下一步加上原来的同等映射x，即F(x)+x
   
   

1.2.7 感受野：越大越好

二、各层的使用

2.1 容器Containers：nn.Module

import torch
from torch import nn
class Mywork(nn.Module):
    # 可以使用alt+insert重写/实现方法
    def __init__(self):
        super().__init__()

    # 前向传播
    def forward(self, input):
        output = input + 1
        return output
mywork = Mywork()
x = torch.tensor(1.0)
output = mywork(x)
print(output)
输出：
tensor(2.)

2.2 卷积层Convolution Layers：nn.conv2d

1. 代码演示

import torch
import torchvision.datasets
from torch import nn
from torch.nn import Conv2d
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

# 演示卷积层
# 进过卷积操作：64×3×32×32-》64×6×30×30
# 最后因为6通道无法输出，需要变成3通道：64×6×30×30-》128×3×30×30

dataset = torchvision.datasets.CIFAR10("./dataset", train=False,
                                       transform=torchvision.transforms.ToTensor(), download=True)
# 一次取64张图片
dataloader = DataLoader(dataset, batch_size=64)


class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        # 卷积设置：输入通道为3，输出通道为6，卷积核大小3×3，步数为1，不填充
        self.conv1 = Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=1, padding=0)

    def forward(self, x):
        x = self.conv1(x)
        return x


mydata = Mydata()
step = 0
writer = SummaryWriter("logs")
for data in dataloader:
    imgs, targets = data
    output = mydata(imgs)
    # torch.Size([64, 3, 32, 32])
    # torch.Size([64, 6, 30, 30])
    # 可知输入64×3×32×32-》64×6×32×32  （数量，通道，图长，图宽）
    print(imgs.shape)
    print(output.shape)

    writer.add_images("inputImage", imgs, step)
    # 这里64×6×30×30无法直接输出，需要变成 yyy×3×30×30
    # 下面参数-1表示未知，经过计算可得出为128
    output = torch.reshape(output, (-1, 3, 30, 30))
    writer.add_images("outputImage", output, step)
    step = step + 1

writer.close()

2. 结果展示：可以看出经过卷积操作图片的变化

2.3 填充层Padding Layers

2.4 池化层Pooling Layers

1. 代码演示

import torchvision.datasets
from torch import nn
from torch.nn import MaxPool2d
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

dataset = torchvision.datasets.CIFAR10("./dataset", train=False,
                                       transform=torchvision.transforms.ToTensor(), download=True)
# 一次取64张图片
dataloader = DataLoader(dataset, batch_size=64)

class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        # ceil_mode表示池化核移动超出边界是否保留
        self.maxpool1 = MaxPool2d(kernel_size=3, ceil_mode=True)

    def forward(self, x):
        x = self.maxpool1(x)
        return x

mydata = Mydata()
step = 0
writer = SummaryWriter("logs")
for data in dataloader:
    imgs, targets = data
    output = mydata(imgs)
    writer.add_images("polling_inputImage", imgs, step)
    writer.add_images("polling_outputImage", output, step)
    step = step + 1

writer.close()

2. 结果展示
   

2.5 非线性激活Non-linear Activations

1. 代码展示

import torchvision.datasets
from torch import nn
from torch.nn import ReLU, Sigmoid
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

dataset = torchvision.datasets.CIFAR10("./dataset", train=False,
                                       transform=torchvision.transforms.ToTensor(), download=True)
# 一次取64张图片
dataloader = DataLoader(dataset, batch_size=64)

class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        self.relu1 = ReLU()
        self.sigmoid1 = Sigmoid()

    def forward(self, x):
        x = self.sigmoid1(x)
        return x

mydata = Mydata()
step = 0
writer = SummaryWriter("logs")
for data in dataloader:
    imgs, targets = data
    output = mydata(imgs)
    writer.add_images("relu_inputImage", imgs, step)
    writer.add_images("relu_outputImage", output, step)
    step = step + 1

writer.close()

2. 结果展示
   

2.6 线性层linear Layers

通过线性层，给定输入x1…xn得到g1…gL

1. 代码演示

import torch
import torchvision.datasets
from torch import nn
from torch.nn import Linear
from torch.utils.data import DataLoader

dataset = torchvision.datasets.CIFAR10("./dataset", train=False,
                                       transform=torchvision.transforms.ToTensor(), download=True)
# 一次取64张图片
dataloader = DataLoader(dataset, batch_size=64)


class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        self.linear1 = Linear(196608, 10)

    def forward(self, x):
        x = self.linear1(x)
        return x

# [64, 3, 32, 32]->[1, 1, 1, 196608]通过线性层将196608->10
mydata = Mydata()
for data in dataloader:
    imgs, targets = data
    print(imgs.shape)  # [64, 3, 32, 32]
    output = torch.reshape(imgs, (1, 1, 1, -1))
    # output = torch.flatten(imgs)
    print(output.shape)  # [1, 1, 1, 196608]
    output = mydata(output)
    print(output.shape)  # [1, 1, 1, 10]

2. 结果展示

torch.Size([64, 3, 32, 32])
torch.Size([1, 1, 1, 196608])
torch.Size([1, 1, 1, 10])

2.7 Sequential层：实现序列化模块

Sequential层的作用：将以上各层封装起来，简化操作
以下实现CIFAR10网络模型

1. 代码演示

import torch
from torch import nn
from torch.nn import Conv2d, MaxPool2d, Flatten, Linear, Sequential


class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        self.model1 = Sequential(
            Conv2d(3, 32, 5, padding=2),
            MaxPool2d(2),
            Conv2d(32, 32, 5, padding=2),
            MaxPool2d(2),
            Conv2d(32, 64, 5, padding=2),
            MaxPool2d(2),
            Flatten(),
            Linear(1024, 64),  # 64*4*4=1024
            Linear(64, 10)
        )

    def forward(self, x):
        x = self.model1(x)
        return x


mydata = Mydata()
print(mydata)
input = torch.ones((64, 3, 32, 32))
output = mydata(input)
print(output.shape)

2. 结果

Mydata(
  (model1): Sequential(
    (0): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (2): Conv2d(32, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (4): Conv2d(32, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (6): Flatten(start_dim=1, end_dim=-1)
    (7): Linear(in_features=1024, out_features=64, bias=True)
    (8): Linear(in_features=64, out_features=10, bias=True)
  )
)
torch.Size([64, 10])

2.8 Loss Funcion

计算实际输出与目标输出的差距，为我们更新输出提供一定依据（反向传播），然后反向传播会计算梯度，最后利用优化器更新参数；持续上述，每次开始时梯度要清零。

1. 代码演示

import torch
import torchvision
from torch import nn
from torch.nn import L1Loss, MSELoss, Conv2d, MaxPool2d, Flatten, Linear, Sequential
from torch.utils.data import DataLoader

inputs = torch.tensor([1, 2, 3], dtype=torch.float32)
targets = torch.tensor([1, 2, 5], dtype=torch.float32)

inputs = torch.reshape(inputs, (1, 1, 1, 3))
targets = torch.reshape(targets, (1, 1, 1, 3))

# Loss函数L1Loss
loss1 = L1Loss()
result1 = loss1(inputs, targets)
print(result1)
# Loss函数MSELoss
loss2 = MSELoss()
result2 = loss2(inputs, targets)
print(result2)
# Loss函数CrossEntropyLoss
x = torch.tensor([0.1, 0.2, 0.3])  # x表示各个类的概率
y = torch.tensor([1])
x = torch.reshape(x, [1, 3])  # 这个3表示3个类
loss3 = nn.CrossEntropyLoss()
result3 = loss3(x, y)
print(result3)

# 演示
dataset = torchvision.datasets.CIFAR10("./dataset", train=False,
                                       transform=torchvision.transforms.ToTensor(), download=True)
# 一次取64张图片
dataloader = DataLoader(dataset, batch_size=1)
class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        self.model1 = Sequential(
            Conv2d(3, 32, 5, padding=2),
            MaxPool2d(2),
            Conv2d(32, 32, 5, padding=2),
            MaxPool2d(2),
            Conv2d(32, 64, 5, padding=2),
            MaxPool2d(2),
            Flatten(),
            Linear(1024, 64),  # 64*4*4=1024
            Linear(64, 10)
        )

    def forward(self, x):
        x = self.model1(x)
        return x

loss = nn.CrossEntropyLoss()
mydata = Mydata()
for data in dataloader:
    imgs, targets = data
    output = mydata(imgs)
    # 计算loss
    result_loss = loss(output, targets)
    print(result_loss)
    # 为反向传播提供依据
    result_loss.backward()

2. 结果

tensor(0.6667)
tensor(1.3333)
tensor(1.1019)
Files already downloaded and verified
tensor(2.2537, grad_fn=<NllLossBackward0>)
...省略

2.9 优化器optim

每次开始时梯度要清零，然后计算loss，然后反向传播会计算梯度，最后利用优化器更新参数。

1. 代码演示

import torch
import torchvision
from torch import nn
from torch.nn import L1Loss, MSELoss, Conv2d, MaxPool2d, Flatten, Linear, Sequential
from torch.utils.data import DataLoader


dataset = torchvision.datasets.CIFAR10("./dataset", train=False,
                                       transform=torchvision.transforms.ToTensor(), download=True)
# 一次取64张图片
dataloader = DataLoader(dataset, batch_size=1)
class Mydata(nn.Module):
    def __init__(self):
        super(Mydata, self).__init__()
        self.model1 = Sequential(
            Conv2d(3, 32, 5, padding=2),
            MaxPool2d(2),
            Conv2d(32, 32, 5, padding=2),
            MaxPool2d(2),
            Conv2d(32, 64, 5, padding=2),
            MaxPool2d(2),
            Flatten(),
            Linear(1024, 64),  # 64*4*4=1024
            Linear(64, 10)
        )

    def forward(self, x):
        x = self.model1(x)
        return x


mydata = Mydata()
# 使用CrossEntropyLoss损失函数
loss = nn.CrossEntropyLoss()
# 使用SGD优化器 lr表示学习速率
optim = torch.optim.SGD(mydata.parameters(), lr=0.01)
for epoch in range(20):  # 进行20轮
    running_loss = 0.0
    for data in dataloader:
        imgs, targets = data
        output = mydata(imgs)
        # 计算loss
        result_loss = loss(output, targets)
        # 参数的梯度清零
        optim.zero_grad()
        # 反向传播计算参数的梯度grad
        result_loss.backward()
        # 利用梯度优化参数
        optim.step()
        running_loss = running_loss + result_loss

    print("每一轮的全部数据的误差：")
    print(running_loss)

2. 结果

Files already downloaded and verified
每一轮的全部数据的误差：
tensor(18657.6328, grad_fn=<AddBackward0>)
...省略