CNN --入门MNIST识别

Smiling & Weeping

　　　　　　　　　　　　　　　　---- 下次你撑伞低头看水洼，

　　　　　　　　　　　　　　　　　　就会想起我说雨是神的烟花。

简介：主要是看刘二大人的视频讲解：https://www.bilibili.com/video/BV1Y7411d7Ys/?spm_id_from=333.337.search-card.all.click

题目及提交链接:Digit Recognizer | Kaggle

深度学习入门的学习项目，使用CNN(Convolutional Nerual Network)

对于Basic CNN的理解：

1. 分成两个部分：前一个部分叫做Feature Extraction,后一部分叫做Classification(其中Feature Extraction又可以分为Convolution,Subsampling等)
2. 其中要求卷积核的通道数量与输入通道数量一致。这种卷积核的总数和输出通道数目的总数一致（详见链接PDF）
3. 卷积(convolution)后,C(channels)，W(width)，H(height)，其中padding和pooling（小技巧：若要卷积W,H不变，取整kernel_size/2）
4. 卷积层：保存图像的空间信息
5. 卷积层要求输入输出是四维张量（B,C,W,H），全连接层的输入输出都是二维张量（B,Input_feature）
6. 卷积（线性变换），激活函数（非线性变换），池化；这个过程若干次后，view打平，进入全连接层

  1 import torch
  2 import torch.nn.functional as F
  3 import torch.nn as nn
  4 import torch.optim as optim
  5 import torch.autograd as lr_scheduler
  6 from torch.utils.data import DataLoader, Dataset
  7 from torchvision import transforms
  8 from torchvision.utils import make_grid
  9 from torchvision import datasets
 10 from torch.autograd import Variable
 11 from sklearn.model_selection import train_test_split
 12 import pandas as pd
 13 import numpy as np
 14 import matplotlib.pyplot as plt
 15
 16 batch_size = 64
 17 transform = transforms.Compose([transforms.ToTensor()])
 18 train_dataset = datasets.MNIST(root='../dataset/mnist/', train=True, download=True, transform=transform)
 19 train_loader = torch.utils.data.DataLoader(dataset=train_dataset, shuffle=True, batch_size=batch_size)
 20 #同样的方式加载一下测试集
 21 test_dataset = datasets.MNIST(root='../dataset/mnist/',  train=False, download=True, transform=transform)
 22 test_loader = torch.utils.data.DataLoader(dataset=test_dataset, shuffle=False,  batch_size=batch_size)
 23
 24 # 使用卷积神经网络进行图像特征提取
 25 # (batch, 1, 28, 28) -> (batch, 10, 24, 24) -> 池化 (batch, 10, 12, 12) -> (batch, 20, 8, 8) -> (batch, 20 , 4, 4) -> (batch, 320) -> (batch, 10)
 26 class Net(torch.nn.Module):
 27     def __init__(self):
 28         super(Net, self).__init__()
 29         self.conv1 = torch.nn.Conv2d(1, 10, kernel_size = 5)
 30         self.conv2 = torch.nn.Conv2d(10, 20, kernel_size = 5)
 31         self.pooling = torch.nn.MaxPool2d(2)
 32         self.fc = torch.nn.Linear(320, 10)
 33
 34     def forward(self, x):
 35         # Flatten data from (n, 1, 28, 28) to (n, 784)
 36         batch_size = x.size(0)
 37         x = F.relu(self.pooling(self.conv1(x)))
 38         x = F.relu(self.pooling(self.conv2(x)))
 39         x = x.view(batch_size, -1) # Flatten
 40         x = self.fc(x)
 41         return x
 42
 43 model = Net()
 44 # print(model)
 45 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 46 model.to(device)
 47 criterion = torch.nn.CrossEntropyLoss(size_average=True)
 48 optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
 49
 50 def train(epoch):
 51     running_loss = 0.0
 52     for batch_idx, data in enumerate(train_loader, 0):
 53         inputs, target = data
 54         inputs, target = inputs.to(device), target.to(device)
 55         optimizer.zero_grad()
 56
 57         # forward + backward + update
 58         outputs = model(inputs)
 59         # 计算真实值 和 测量值 之间的误差
 60         loss = criterion(outputs, target)
 61         loss.backward()
 62         optimizer.step()
 63
 64         running_loss += loss.item()
 65         if batch_idx % 300 == 299:
 66             print('[%d, %5d] loss: %3f' % (epoch + 1, batch_idx+1, running_loss / 2000))
 67             running_loss = 0.0
 68
 69 def test():
 70     correct = 0
 71     total = 0
 72     with torch.no_grad():
 73         for data in test_loader:
 74             inputs, target = data
 75             inputs, target = inputs.to(device), target.to(device)
 76             outputs = model(inputs)
 77             _, prediction = torch.max(outputs.data, dim=1)
 78             total += target.size(0)
 79             correct += (prediction == target).sum().item()
 80     print('Accuracy on test set: %d %% [%d/%d]' % (100*correct / total, correct, total))
 81     return correct/total
 82
 83 epoch_list = []
 84 acc_list = []
 85 for epoch in range(10):
 86     train(epoch)
 87     acc = test()
 88     epoch_list.append(epoch)
 89     acc_list.append(acc)
 90
 91 plt.plot(epoch_list, acc_list)
 92 plt.ylabel("accuracy")
 93 plt.xlabel("epoch")
 94 plt.show()
 95
 96 class DatasetSubmissionMNIST(torch.utils.data.Dataset):
 97     def __init__(self, file_path, transform=None):
 98         self.data = pd.read_csv(file_path)
 99         self.transform = transform
100
101     def __len__(self):
102         return len(self.data)
103
104     def __getitem__(self, index):
105         image = self.data.iloc[index].values.astype(np.uint8).reshape((28, 28, 1))
106
107
108         if self.transform is not None:
109             image = self.transform(image)
110
111         return image
112
113 transform = transforms.Compose([
114     transforms.ToPILImage(),
115     transforms.ToTensor(),
116     transforms.Normalize(mean=(0.5,), std=(0.5,))
117 ])
118
119 submissionset = DatasetSubmissionMNIST('/kaggle/input/digit-recognizer/test.csv', transform=transform)
120 submissionloader = torch.utils.data.DataLoader(submissionset, batch_size=batch_size, shuffle=False)
121
122 submission = [['ImageId', 'Label']]
123
124 with torch.no_grad():
125     model.eval()
126     image_id = 1
127
128     for images in submissionloader:
129         images = images.cuda()
130         log_ps = model(images)
131         ps = torch.exp(log_ps)
132         top_p, top_class = ps.topk(1, dim=1)
133
134         for prediction in top_class:
135             submission.append([image_id, prediction.item()])
136             image_id += 1
137
138 print(len(submission) - 1)
139 import csv
140
141 with open('submission.csv', 'w') as submissionFile:
142     writer = csv.writer(submissionFile)
143     writer.writerows(submission)
144
145 print('Submission Complete!')
146 # summission.to_csv('/kaggle/working/submission.csv', index=False)

就效果来说，也就一般，后面的Advance CNN 会有更高的效率和准确性，大家可以敲一下代码放在自己的编译器上跑一下

对了，这是GPU版本，若用CPU，把所有的device删除就可以，--<-<-<@

文章到此结束，我们下次再见

一束光线，可能会摔碎

　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　   但仍旧光芒四射