赞
踩
今天是尝试用 PyTorch 框架来跑 MNIST 手写数字数据集的第一天,主要学习 LeNet 网络结构的定义。本 blog 主要记录一个学习的路径以及学习资料的汇总。
注意:这是用 Python 2.7 版本写的代码
第一天(LeNet 网络的搭建):https://blog.csdn.net/qq_36627158/article/details/108098147
第二天(加载 MNIST 数据集):https://blog.csdn.net/qq_36627158/article/details/108119048
第三天(训练模型):https://blog.csdn.net/qq_36627158/article/details/108163693
第四天(单例测试):https://blog.csdn.net/qq_36627158/article/details/108183655
感谢 凯神 提供的代码与耐心指导!
- import torch.nn as nn
- import torch.nn.functional as F
-
-
- class Net(nn.Module):
- def __init__(self):
- super(Net, self).__init__()
-
- # C1: input: 1*28*28 output: 6*28*28 (32-5+1)
- self.conv1 = nn.Conv2d(
- in_channels=1,
- out_channels=6,
- kernel_size=5,
- padding=2
- )
-
- # S2: input: 6*28*28 output: 6*14*14
- self.max_pooling1 = nn.MaxPool2d(
- kernel_size=2,
- stride=2
- )
-
- # C3: input: 6*14*14 output: 16*10*10 (14-5+1)
- self.conv2 = nn.Conv2d(
- in_channels=6,
- out_channels=16,
- kernel_size=5
- )
-
- # S4: input: 16*10*10 output: 16*5*5
- self.max_pooling2 = nn.MaxPool2d(
- kernel_size=2,
- stride=2
- )
-
- # C5: input: 16*5*5 output: 1*120*1
- self.conv3 = nn.Linear(
- in_features=16 * 5 * 5,
- out_features=120
- )
-
- # F6: input: 1*120*1 output: 1*84*1
- self.fc1 = nn.Linear(
- in_features=120,
- out_features=84
- )
-
- # F7(OutPut): input: 1*84*1 output: 1*10*1
- self.fc2 = nn.Linear(
- in_features=84,
- out_features=10
- )
-
- def forward(self, x):
- # C1:
- x = self.conv1(x)
- x = F.relu(x)
-
- # S2:
- x = self.max_pooling1(x)
-
- # C3:
- x = self.conv2(x)
- x = F.relu(x)
-
- # S4:
- x = self.max_pooling2(x)
-
- # num_flat_features(x): get a flatten vector's size
- # view(): 16*5*5 -> 400*1
- x = x.view(-1, self.num_flat_features(x))
-
- # C5:
- x = self.conv3(x)
- x = F.relu(x)
-
- # F6:
- x = self.fc1(x)
- x = F.relu(x)
-
- # F7:
- x = self.fc2(x)
- x = F.relu(x)
-
- return x
-
- def num_flat_features(self, x):
- # x.size() = [b, c, h, w]
- # size = [c, h, w] [16*5*5]
- size = x.size()[1:]
-
- # c * h * w
- num_features = 1
- for s in size:
- num_features *= s
- return num_features
https://blog.csdn.net/qq_42570457/article/details/81460807
https://www.kaggle.com/jidhumohan/mnist-png?
https://www.runoob.com/python3/python3-class.html
https://www.runoob.com/python/python-func-super.html
最开始是因为看到了池化层既可以写成:self.max_pooling1 = nn.MaxPool2d() ;也可以写成:self.max_pooling1 = F.maxpool2d()。就产生了疑问:既然都有每一层的实现方法,为什么不只使用 torch.nn ?或者为什么不只使用 torch.nn.functional ?为什么要两个混在一起写?有什么区别吗?
后来找到了一篇不错的解释:https://blog.csdn.net/GZHermit/article/details/78730856
表示去掉列表中第一个元素(下标为0),对后面的元素进行操作。
https://blog.csdn.net/gaofengyan/article/details/90697743
https://blog.csdn.net/zkq_1986/article/details/100319146
四维指的是:【b,c,h,w】:b:图片的张数 c:通道数 h:高 w:宽
一开始以为代码里的forward() 函数是自己写的 /(ㄒoㄒ)/~~ 不对!是 重写 父类 nn.Module 里的 forward() 函数
注意:定义神经网络, 需要继承 nn.Moudle, 并重载 __init__ 和 forward 方法
forward() 函数使用我们在构造函数(初始化函数)内部定义的所有层,其实就是实际的网络转换
https://cloud.tencent.com/developer/article/1639430
forward 的使用:https://blog.csdn.net/xu380393916/article/details/97280035
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。