使用PyTorch实现CNN

使用PyTorch实现CNN

1. 导入所需包:

import torch 
from torch.utils import data # 获取迭代数据
from torch.autograd import Variable # 获取变量
import torchvision
from torchvision.datasets import mnist # 获取数据集
import matplotlib.pyplot as plt

2. 获取数据集

2.1 获取数据集,并对数据集进行预处理

(1)对原有数据转成Tensor类型

(2)用平均值和标准偏差归一化张量图像

# 数据集的预处理
data_tf = torchvision.transforms.Compose(
    [
        torchvision.transforms.ToTensor(),
        torchvision.transforms.Normalize([0.5],[0.5])
    ]
)

data_path = r'C:\Users\liev\Desktop\myproject\yin_test\MNIST_DATA_PyTorch'
# 获取数据集
train_data = mnist.MNIST(data_path,train=True,transform=data_tf,download=False)
test_data = mnist.MNIST(data_path,train=False,transform=data_tf,download=False)

第一次下载的输出:

Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
Processing...
Done!

注意:

  1. 对数据的预处理还有很多。
  2. 第一次获取数据集时,参数download=True,会下载MNIST数据集所有文件,包括训练集和测试集
  3. 获取MNIST数据集的步骤:
  • 如果本地没有数据集:

    • train_data = mnist.MNIST(data_path,train=True,transform=data_tf,download=True)
      
    • 等待下载,直到下载完成

    • train_data = mnist.MNIST(data_path,train=True,transform=data_tf,download=False)
      test_data = mnist.MNIST(data_path,train=False,transform=data_tf,download=False)
      
    • 获取测试集和训练集

  • 如果本地有数据集

    • train_data = mnist.MNIST(data_path,train=True,transform=data_tf,download=False)
      test_data = mnist.MNIST(data_path,train=False,transform=data_tf,download=False)
      

2.2 获取迭代数据:data.DataLoader()

train_loader = data.DataLoader(train_data,batch_size=128,shuffle=True)
test_loader = data.DataLoader(test_data,batch_size=100,shuffle=True)

注意:

  1. DataLoader返回的是所有的数据,只是分成了每批次为参数batch_size的数据
  2. DataLoader的shuffle参数,True 决定了是否能多次取出batch_size,False,则表明只能取出数据集大小的数据。

3. 定义网络结构

CNN网络结构输入shape卷积核激活函数输出图像
conv1[128,1,28,28][3,3,1,16]ReLU[128, 16, 14, 14]
conv2[128, 16, 14, 14][3,3,16,32]ReLU[128, 32, 7, 7]
conv3[128, 32, 7, 7][3,3,32,64]ReLU[128, 64, 4, 4]
conv4[128, 64, 4, 4][3,3,64,64]ReLU[128, 64, 2, 2]

代码实现:

# 定义网络结构
class CNNnet(torch.nn.Module):
    def __init__(self):
        super(CNNnet,self).__init__()
        self.conv1 = torch.nn.Sequential(
            torch.nn.Conv2d(in_channels=1,
                            out_channels=16,
                            kernel_size=3,
                            stride=2,
                            padding=1),
            torch.nn.BatchNorm2d(16),
            torch.nn.ReLU()
        )
        self.conv2 = torch.nn.Sequential(
            torch.nn.Conv2d(16,32,3,2,1),
            torch.nn.BatchNorm2d(32),
            torch.nn.ReLU()
        )
        self.conv3 = torch.nn.Sequential(
            torch.nn.Conv2d(32,64,3,2,1),
            torch.nn.BatchNorm2d(64),
            torch.nn.ReLU()
        )
        self.conv4 = torch.nn.Sequential(
            torch.nn.Conv2d(64,64,2,2,0),
            torch.nn.BatchNorm2d(64),
            torch.nn.ReLU()
        )
        self.mlp1 = torch.nn.Linear(2*2*64,100)
        self.mlp2 = torch.nn.Linear(100,10)
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        x = self.mlp1(x.view(x.size(0),-1))
        x = self.mlp2(x)
        return x
model = CNNnet()
print(model)

输出:

CNNnet(
  (conv1): Sequential(
    (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv2): Sequential(
    (0): Conv2d(16, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv3): Sequential(
    (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv4): Sequential(
    (0): Conv2d(64, 64, kernel_size=(2, 2), stride=(2, 2))
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (mlp1): Linear(in_features=256, out_features=100, bias=True)
  (mlp2): Linear(in_features=100, out_features=10, bias=True)
)

4. 定义损失和优化器

(1)使用交叉熵损失

(2)使用Adam优化器

loss_func = torch.nn.CrossEntropyLoss()
opt = torch.optim.Adam(model.parameters(),lr=0.001)
  1. model.parmaters()含义:

  2. 使用损失和优化器的步骤:

    • 获取损失:loss = loss_func(out,batch_y)
    • 清空上一步残余更新参数:opt.zero_grad()
    • 误差反向传播:loss.backward()
    • 将参数更新值施加到netparmeters上:opt.step()

5. 训练网络

loss_count = []
for epoch in range(2):
    for i,(x,y) in enumerate(train_loader):
        batch_x = Variable(x) # torch.Size([128, 1, 28, 28])
        batch_y = Variable(y) # torch.Size([128])
        # 获取最后输出
        out = model(batch_x) # torch.Size([128,10])
        # 获取损失
        loss = loss_func(out,batch_y)
        # 使用优化器优化损失
        opt.zero_grad()  # 清空上一步残余更新参数值
        loss.backward() # 误差反向传播,计算参数更新值
        opt.step() # 将参数更新值施加到net的parmeters上
        if i%20 == 0:
            loss_count.append(loss)
            print('{}:\t'.format(i), loss.item())
            torch.save(model,r'C:\Users\liev\Desktop\myproject\yin_test\log_CNN')
        if i % 100 == 0:
            for a,b in test_loader:
                test_x = Variable(a)
                test_y = Variable(b)
                out = model(test_x)
                # print('test_out:\t',torch.max(out,1)[1])
                # print('test_y:\t',test_y)
                accuracy = torch.max(out,1)[1].numpy() == test_y.numpy()
                print('accuracy:\t',accuracy.mean())
                break
plt.figure('PyTorch_CNN_Loss')
plt.plot(loss_count,label='Loss')
plt.legend()
plt.show()

输出:

0:	 2.313704252243042
accuracy:	 0.11
20:	 1.1835652589797974
40:	 0.5378416776657104
60:	 0.41440480947494507
80:	 0.18270650506019592
100:	 0.18721994757652283
accuracy:	 0.92
......
380:	 0.032591354101896286
400:	 0.024792633950710297
accuracy:	 1.0
420:	 0.03427279368042946
440:	 0.04764523729681969
460:	 0.01753203198313713

损失图:

PyTorch_CNN_CrossEntropyLoss.png

如果使用MSELoss:平方差损失

将真实值转为one-hot形式

def one_hot(data):
    hot = np.zeros([10])
    hot[data] = 1
    return hot
# 并且在计算损失前,加入下面一行代码,将真实值转为one-hot形式
y = [one_hot(i) for i in y]

7. 测试网络:精确度:0.98

注意点:

  1. 获取accuracy时的问题:
    • 获取的预测值shape:【128,10】
    • 真实值shape:【128】,无one-hot编码
    • 转换预测值:使用torch.max() 获取最后输出的每张图片预测值的最大值
# 测试网络
model = torch.load(r'C:\Users\liev\Desktop\myproject\yin_test\log_CNN')
accuracy_sum = []
for i,(test_x,test_y) in enumerate(test_loader):
    test_x = Variable(test_x)
    test_y = Variable(test_y)
    out = model(test_x)
    # print('test_out:\t',torch.max(out,1)[1])
    # print('test_y:\t',test_y)
    accuracy = torch.max(out,1)[1].numpy() == test_y.numpy()
    accuracy_sum.append(accuracy.mean())
    print('accuracy:\t',accuracy.mean())

print('总准确率:\t',sum(accuracy_sum)/len(accuracy_sum))
# 精确率图
print('总准确率:\t',sum(accuracy_sum)/len(accuracy_sum))
plt.figure('Accuracy')
plt.plot(accuracy_sum,'o',label='accuracy')
plt.title('Pytorch_CNN_Accuracy')
plt.legend()
plt.show()

输出:

accuracy:	 0.98
accuracy:	 1.0
accuracy:	 1.0
accuracy:	 1.
总准确率:	 0.9850999999999999

8. 其他实验:

8.1 全连接第一层增加ReLU激活函数:提高了0.02

测试输出:

accuracy:	 0.98
accuracy:	 0.99
accuracy:	 0.98
accuracy:	 0.99
总准确率:	 0.9872999999999992

8.2 去掉批量归一化:降低了0.01

accuracy:	 0.97
accuracy:	 0.97
accuracy:	 0.92
总准确率:	 0.9746999999999996

8.3 使用LeakyReLU激活函数:降低0.01

accuracy:	 0.97
accuracy:	 0.98
accuracy:	 1.0
总准确率:	 0.9848999999999997

8.4 使用PReLU激活函数:提升0.01

accuracy:	 0.97
accuracy:	 1.0
accuracy:	 1.0
accuracy:	 0.97
总准确率:	 0.9867999999999998

卷积方法的使用:torch.nn.Conv2d()

在由多个输入平面组成的输入信号上应用2D卷积。

在最简单的情况下,具有输入大小的图层的输出值
( N , C i n , H , W ) ( N , C i n , H , W ) (N,C_{in},H,W)(N,Cin,H,W) (N,Cin,H,W)(N,Cin,H,W)
和输出
( N , C o u t , H o u t , W o u t ) ( N , C o u t , H o u t , W o u t ) (N,Cout,Hout,Wout)(N,Cout,Hout,Wout) (N,Cout,Hout,Wout)(N,Cout,Hout,Wout)
可以精确地描述为:

out ( N i , C o u t j ) = bias ( C o u t j ) + ∑ k = 0 C i n − 1 weight ( C o u t j , k ) ⋆ input ( N i , k ) \text{out}(N_i,C_{out_j})=\text{bias}(C_{out_j}) + \sum_{k = 0}^{C_{in}-1}\text{weight}(C_{out_j}, k) \star\text{input}(N_i, k) out(Ni,Coutj)=bias(Coutj)+k=0Cin1weight(Coutj,k)input(Ni,k)

⋆是有效的2D 互相关运算符, N 是一个批量大小, C 表示多个频道, H 是输入平面的高度,以像素为单位 W 是像素的宽度。

参数说明:

参数说明
in_channels输入图像中的通道数:int
out_channels卷积产生的通道数
kernel_size卷积内核的大小
stride卷积的步幅。默认值:1
padding将零填充添加到输入的两侧。默认值:0
dilation内核元素之间的间距。默认值:1
groups从输入通道到输出通道的阻塞连接数。默认值:1
bias如果True,在输出中增加了可学习的偏差。默认:True

计算输出图片shape:

  • 输入shape格式:

( N , C i n , H i n , W i n ) (N,C_{in},H_{in},W_{in}) (N,Cin,Hin,Win)

  • 输出shape格式:

( N , C o u t , H o u t , W o u t ) (N,C_{out},H_{out},W_{out}) (N,Cout,Hout,Wout)

输出图片shape的计算公式:

H o u t = ⌊ H i n + 2 × p a d d i n g [ 0 ] − d i l a t i o n [ 0 ] × ( k e r n e l _ s i z e [ 0 ] − 1 ) − 1 s t r i d e [ 0 ] + 1 ⌋ W o u t = ⌊ W i n + 2 × p a d d i n g [ 1 ] − d i l a t i o n [ 1 ] × ( k e r n e l _ s i z e [ 1 ] − 1 ) − 1 s t r i d e [ 1 ] + 1 ⌋ H_{out} = \bigg\lfloor\frac{\mathbf{H}_{\mathbf{in}}+2\times \mathbf{padding[0]}-\mathbf{dilation[0]}\times (\mathbf{kernel}\_\mathbf{size[0]}-1)-1 }{\mathbf{stride[0]}}+1 \bigg\rfloor \\ W_{out} = \bigg\lfloor\frac{\mathbf{W}_{\mathbf{in}}+2\times \mathbf{padding[1]}-\mathbf{dilation[1]}\times (\mathbf{kernel}\_\mathbf{size[1]}-1)-1 }{\mathbf{stride[1]}}+1 \bigg\rfloor Hout=stride[0]Hin+2×padding[0]dilation[0]×(kernel_size[0]1)1+1Wout=stride[1]Win+2×padding[1]dilation[1]×(kernel_size[1]1)1+1

变量:

  • weight(Tensor) - 形状模块的可学习权重(out_channels,in_channels,kernel_size [0],kernel_size [1])
  • 偏见(Tensor) - 形状模块的可学习偏差(out_channels)

实例代码:

import torch

conv = torch.nn.Conv2d(1,3,2,1,0)

print('conv.weight.size():\t',conv.weight.size())
print('conv.bias.size():\t',conv.bias.size())
print('conv初始化的weight数据:\n',conv.weight)
print('conv初始化的bias数据:\n',conv.bias)

输出:

conv.weight.size():	 torch.Size([3, 1, 2, 2])
conv.bias.size():	 torch.Size([3])
conv初始化的weight数据:
 Parameter containing:
tensor([[[[ 0.2753, -0.1573],
          [-0.4197,  0.1795]]],


        [[[ 0.1529,  0.3869],
          [ 0.0219, -0.2829]]],


        [[[ 0.3727, -0.1673],
          [ 0.4687,  0.3792]]]], requires_grad=True)
conv初始化的bias数据:
 Parameter containing:
tensor([ 0.2909, -0.0980,  0.0066], requires_grad=True)

反卷积方法的使用:torch.nn.ConvTranspose2d()

在由多个输入平面组成的输入图像上应用2D转置卷积运算符。

该模块可以看作Conv2d相对于其输入的梯度。它也被称为分数跨度卷积或反卷积(尽管它不是实际的去卷积操作)。

参数说明:

参数说明
in_channels输入图像中的通道数
out_channels卷积产生的通道数
kernel_size卷积内核的大小
stride卷积的步幅。默认值:1
padding零填充将添加到输入中每个维度的两侧。默认值:0kernel_size - 1 - padding
output_padding在输出形状中添加到每个维度的一侧的附加大小。默认值:0
groups从输入通道到输出通道的阻塞连接数。默认值:1
bias如果True,在输出中增加了可学习的偏差。默认:True
dilation内核元素之间的间距。默认值:1

计算输出图片shape:

  • 输入:

( N , C i n , H i n , W i n ) (N,C_{in},H_{in},W_{in}) (N,Cin,Hin,Win)

  • 输出:

( N , C o u t , H o u t , W o u t ) (N,C_{out},H_{out},W_{out}) (N,Cout,Hout,Wout)

输出图片shape的计算公式:
H o u t = ( H i n − 1 ) × s t r i d e [ 0 ] − 2 × p a d d i n g [ 0 ] + k e r n e l _ s i z e [ 0 ] + o u t p u t _ p a d d i n g [ 0 ] W o u t = ( W i n − 1 ) × s t r i d e [ 1 ] − 2 × p a d d i n g [ 1 ] + k e r n e l _ s i z e [ 1 ] + o u t p u t _ p a d d i n g [ 1 ] \mathbf{H_{out}} = \mathbf{(H_{in}-1)}\times \mathbf{stride[0]} - 2\times \mathbf{padding[0] }+\mathbf{kernel}\_\mathbf{size[0]}+\mathbf{output}\_\mathbf{padding[0]} \\ \mathbf{W_{out}} = \mathbf{(W_{in}-1)}\times \mathbf{stride[1]} - 2\times \mathbf{padding[1] }+\mathbf{kernel}\_\mathbf{size[1]}+\mathbf{output}\_\mathbf{padding[1]} Hout=(Hin1)×stride[0]2×padding[0]+kernel_size[0]+output_padding[0]Wout=(Win1)×stride[1]2×padding[1]+kernel_size[1]+output_padding[1]

变量:

  • weight(Tensor) - 形状模块的可学习权重(out_channels,in_channels,kernel_size [0],kernel_size [1])
  • 偏见(Tensor) - 形状模块的可学习偏差(out_channels)
©️2020 CSDN 皮肤主题: 大白 设计师:CSDN官方博客 返回首页