pytorch深度學習的套路都差不多，多看多想多寫多測試，自然就會了。主要的技術還是在于背后的數學思想和數學邏輯。
廢話不多說，上代碼自己看。
import torch
import numpy as np
import torch.nn as nn
import torchvision
import torchvision.transforms as transformsclass Network(nn.Module):def __init__(self ,input_dim ,hidden_dim ,out_dim):super().__init__()self.layer1 = nn.Sequential( # 全連接層 [1, 28, 28]nn.Linear(784, 400), # 輸入維度，輸出維度nn.BatchNorm1d(400), # 批標準化，加快收斂，可不需要nn.ReLU() # 激活函數)self.layer2 = nn.Sequential(nn.Linear(400, 200),nn.BatchNorm1d(200),nn.ReLU())self.layer3 = nn.Sequential( # 全連接層nn.Linear(200, 100),nn.BatchNorm1d(100),nn.ReLU())self.layer4 = nn.Sequential( # 最后一層為實際輸出，不需要激活函數，因為有 10 個數字，所以輸出維度為 10，表示10 類nn.Linear(100, 10),)def forward(self ,x):x = self.layer1(x)x = self.layer2(x)x = self.layer3(x)output = self.layer4(x)return output
def get_num_correct(preds, labels):return (preds.argmax(dim=1) == labels).sum().item()def dropout(x, keep_prob = 0.5):'''np.random.binomial 當輸入二維數組時，按行按列（每個維度）都是按照給定概率生成1的個數，
比如 輸入 10 * 6的矩陣，按照0.5的概率生成1 那么每列都大概會有5個1，每行大概會有3個1，
其實就不用考慮按行drop或者按列drop，相當于每行生成的mask都是不一樣的，那么矩陣中每行的元素（代表一層中的神經元）都是按照不同的mask失活的
當矩陣形狀改變行列代表的意義不一樣時，由于每行每列（各個維度）的1的個數都是按照prob留存的，因此對結果沒有影響。'''mask = torch.from_numpy(np.random.binomial(1,keep_prob,x.shape))return x * mask / keep_probif __name__ == "__main__":train_set = torchvision.datasets.MNIST(root='./data', train=True, download=False, transform=transforms.Compose([transforms.ToTensor()]))test_set = torchvision.datasets.MNIST(root='./data',train=False,download=False,transform=transforms.Compose([transforms.ToTensor()]))train_loader = torch.utils.data.DataLoader(train_set, batch_size=512, shuffle=True)test_loader = torch.utils.data.DataLoader(test_set, batch_size=512, shuffle=True)net = Network(28 * 28, 256, 10)optimizer = torch.optim.SGD(net.parameters(), lr=0.01)criterion = nn.CrossEntropyLoss()epoch = 10for i in range(epoch):train_accur = 0.0train_loss = 0.0for batch in train_loader:images, labels = batch#images, labels = images.to(device), labels.to(device)images = images.squeeze(1).reshape(images.shape[0], -1)preds = net(images)optimizer.zero_grad()loss = criterion(preds, labels)loss.backward()optimizer.step()train_loss += loss.item()train_accur += get_num_correct(preds, labels)print("loss :" + str(train_loss) + "train accur:" + str(train_accur * 1.0 / 60000))global correctwith torch.no_grad():correct = 0for batch in test_loader:images, labels = batch#images, labels = images.to(device), labels.to(device)images = images.squeeze(1).reshape(-1, 784)preds = net(images)preds = preds.argmax(dim=1)correct += (preds == labels).sum()print(correct)print(correct.item() * 1.0 / len(test_set))
)
