pytorch深度学习的套路都差不多，多看多想多写多测试，自然就会了。主要的技术还是在于背后的数学思想和数学逻辑。
废话不多说，上代码自己看。
python
import torch
import numpy as np
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
class Network(nn.Module):
def __init__(self ,input_dim ,hidden_dim ,out_dim):
super().__init__()
self.layer1 = nn.Sequential( # 全连接层 [1, 28, 28]
nn.Linear(784, 400), # 输入维度，输出维度
nn.BatchNorm1d(400), # 批标准化，加快收敛，可不需要
nn.ReLU() # 激活函数
)
self.layer2 = nn.Sequential(
nn.Linear(400, 200),
nn.BatchNorm1d(200),
nn.ReLU()
)
self.layer3 = nn.Sequential( # 全连接层
nn.Linear(200, 100),
nn.BatchNorm1d(100),
nn.ReLU()
)
self.layer4 = nn.Sequential( # 最后一层为实际输出，不需要激活函数，因为有 10 个数字，所以输出维度为 10，表示10 类
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 当输入二维数组时，按行按列（每个维度）都是按照给定概率生成1的个数，
比如 输入 10 * 6的矩阵，按照0.5的概率生成1 那么每列都大概会有5个1，每行大概会有3个1，
其实就不用考虑按行drop或者按列drop，相当于每行生成的mask都是不一样的，那么矩阵中每行的元素（代表一层中的神经元）都是按照不同的mask失活的
当矩阵形状改变行列代表的意义不一样时，由于每行每列（各个维度）的1的个数都是按照prob留存的，因此对结果没有影响。
'''
mask = torch.from_numpy(np.random.binomial(1,keep_prob,x.shape))
return x * mask / keep_prob
if __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 = 10
for i in range(epoch):
train_accur = 0.0
train_loss = 0.0
for 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 correct
with torch.no_grad():
correct = 0
for 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))