首先要熟悉一下怎么使用PyTorch来实现前馈神经网络吧．为了方便理解，我们这里只拿只有一个隐藏层的前馈神经网络来举例：
一个前馈神经网络的源码和注释如下：比较简单，这里就不多介绍了．
1 class NeuralNet(nn.Module): 2 def __init__(self, input_size, hidden_size, num_classes): 3 super(NeuralNet, self).__init__() 4 self.fc1 = nn.Linear(input_size, hidden_size) //输入层 5 self.relu = nn.ReLU() //隐藏网络：elu的功能是将输入的feature的tensor所有的元素中如果小于零的就取零。 6 self.fc2 = nn.Linear(hidden_size, num_classes) //输出层 7 8 def forward(self, x): 9 out = self.fc1(x) 10 out = self.relu(out) 11 out = self.fc2(out) 12 return out
下面要看一下怎么调用和使用前馈神经网络的：为了提高运算效率，要把该网络优先使用GPU来进行运算．这里的输入尺寸和隐藏尺寸要和训练的图片保持一致的．
# Device configuration device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = NeuralNet(input_size, hidden_size, num_classes).to(device)
为了训练网络，都需要定义一个loss function来描述模型对问题的求解精度。loss越小，代表模型的结果和真实值偏差越小，这里使用CrossEntropyLoss()来计算．Adam，这是一种基于一阶梯度来优化随机目标函数的算法。详细的概念和推导我们后续再专门做分析．
criterion = nn.CrossEntropyLoss() //针对单目标分类问题, 结合了 nn.LogSoftmax() 和 nn.NLLLoss() 来计算 loss.
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) //优化器，设置学习的速度和使用的模型接下来就是训练模型了，训练模型这部分是有点绕的，首先我们来看代码，后面再针对各个函数做说明：
1 total_step = len(train_loader) 2 for epoch in range(num_epochs): 3 for i, (images, labels) in enumerate(train_loader): 4 # Move tensors to the configured device 5 images = images.reshape(-1, 28*28).to(device) 6 labels = labels.to(device) 7 8 # Forward pass 9 outputs = model(images) 10 loss = criterion(outputs, labels) 11 12 # Backward and optimize 13 optimizer.zero_grad() //把梯度置零，也就是把loss关于weight的导数变成0. 14 loss.backward() 15 optimizer.step()
训练模型，首先把图片矩阵变换成25*25的矩阵单元．其次，把运算参数绑定到特定设备上．
然后就是网络的前向传播了：
outputs = model(inputs)
然后将输出的outputs和原来导入的labels作为loss函数的输入就可以得到损失了：
loss = criterion(outputs, labels)
计算得到loss后就要回传损失。要注意的是这是在训练的时候才会有的操作，测试时候只有forward过程。
loss.backward()
回传损失过程中会计算梯度，然后需要根据这些梯度更新参数，optimizer.step()就是用来更新参数的。optimizer.step()后，你就可以从optimizer.param_groups[0][‘params’]里面看到各个层的梯度和权值信息。
optimizer.step()
测试这个模型，没有梯度的模型，这样就大大大额减少了内存的使用量和运算效率，这个测试模型，其实只有一个关键的语句就可以预测模型了，那就是：_, predicted = torch.max(outputs.data, 1)．
with torch.no_grad(): correct = 0 total = 0 for images, labels in test_loader: images = images.reshape(-1, 28*28).to(device) labels = labels.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) print(labels.size(0)) correct += (predicted == labels).sum().item()
这里有个问题．训练好的数据怎么和预测联系起来呢？
训练输出的outputs也是torch.autograd.Variable格式，得到输出后（网络的全连接层的输出）还希望能到到模型预测该样本属于哪个类别的信息，这里采用torch.max。torch.max()的第一个输入是tensor格式，所以用outputs.data而不是outputs作为输入；第二个参数1是代表dim的意思，也就是取每一行的最大值，其实就是我们常见的取概率最大的那个index；第三个参数loss也是torch.autograd.Variable格式。
总体源码：
1 import torch 2 import torch.nn as nn 3 import torchvision 4 import torchvision.transforms as transforms 5 6 7 # Device configuration 8 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 9 10 # Hyper-parameters 11 input_size = 784 12 hidden_size = 500 13 num_classes = 10 14 #input_size = 84 15 #hidden_size = 50 16 #num_classes = 2 17 num_epochs = 5 18 batch_size = 100 19 learning_rate = 0.001 20 21 # MNIST dataset 22 train_dataset = torchvision.datasets.MNIST(root='../../data', 23 train=True, 24 transform=transforms.ToTensor(), 25 download=True) 26 27 test_dataset = torchvision.datasets.MNIST(root='../../data', 28 train=False, 29 transform=transforms.ToTensor()) 30 31 # Data loader 32 train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 33 batch_size=batch_size, 34 shuffle=True) 35 36 test_loader = torch.utils.data.DataLoader(dataset=test_dataset, 37 batch_size=batch_size, 38 shuffle=False) 39 40 # Fully connected neural network with one hidden layer 41 class NeuralNet(nn.Module): 42 def __init__(self, input_size, hidden_size, num_classes): 43 super(NeuralNet, self).__init__() 44 self.fc1 = nn.Linear(input_size, hidden_size) 45 self.relu = nn.ReLU() 46 self.fc2 = nn.Linear(hidden_size, num_classes) 47 48 def forward(self, x): 49 out = self.fc1(x) 50 out = self.relu(out) 51 out = self.fc2(out) 52 return out 53 54 model = NeuralNet(input_size, hidden_size, num_classes).to(device) 55 56 # Loss and optimizer 57 criterion = nn.CrossEntropyLoss() 58 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) 59 60 # Train the model 61 total_step = len(train_loader) 62 for epoch in range(num_epochs): 63 for i, (images, labels) in enumerate(train_loader): 64 # Move tensors to the configured device 65 images = images.reshape(-1, 28*28).to(device) 66 labels = labels.to(device) 67 68 # Forward pass 69 outputs = model(images) 70 loss = criterion(outputs, labels) 71 72 # Backward and optimize 73 optimizer.zero_grad() 74 loss.backward() 75 optimizer.step() 76 77 if (i+1) % 100 == 0: 78 print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}' 79 .format(epoch+1, num_epochs, i+1, total_step, loss.item())) 80 # Test the model 81 # In test phase, we don't need to compute gradients (for memory efficiency) 82 with torch.no_grad(): 83 correct = 0 84 total = 0 85 for images, labels in test_loader: 86 images = images.reshape(-1, 28*28).to(device) 87 labels = labels.to(device) 88 outputs = model(images) 89 _, predicted = torch.max(outputs.data, 1) 90 total += labels.size(0) 91 #print(predicted) 92 correct += (predicted == labels).sum().item() 93 94 print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total)) 95 96 # Save the model checkpoint 97 torch.save(model.state_dict(), 'model.ckpt')
每日一言：人之所畏，不可不畏。
参考文档：
１ https://blog.csdn.net/fireflychh/article/details/75516165
２ https://blog.csdn.net/kgzhang/article/details/77479737
