[PyTorch]利用torch.nn实现softmax（在Fashion-MNIST数据集上）

实验要求

1. 利用torch.nn实现softmax在Fashion-MNIST数据集上进行训练和测试
2. 从loss，训练集以及测试集上的准确率等多个角度对结果进行分析

一、加载Fashion-MNIST数据集

mnist_train = torchvision.datasets.FashionMNIST(
    root='~/Datasets/FashionMNIST',
    train=True,
    download=True,
    transform=transforms.ToTensor())  # 将所有数据转换为Tensor
mnist_test = torchvision.datasets.FashionMNIST(
    root='~/Datasets/FashionMNIST',
    train=False,
    download=True,
    transform=transforms.ToTensor()
)

二、通过Dataloader读取小批量数据样本

# 2、通过Dataloader读取小批量数据样本
batch_size = 256
train_iter = torch.utils.data.DataLoader(
    mnist_train,
    batch_size=batch_size,
    shuffle=True,
    num_workers=0
)
test_iter = torch.utils.data.DataLoader(
    mnist_train,
    batch_size=batch_size,
    shuffle=False,
    num_workers=0
)

三、构建模型

num_inputs = 784
num_outputs = 10  # 共10类


# 3、构建模型
class softmaxnet(torch.nn.Module):
    def __init__(self, n_features, n_labels):
        super(softmaxnet, self).__init__()
        self.linear = torch.nn.Linear(n_features, n_labels)

    def softmax(self, X):  # softmax计算
        X_exp = X.exp()  # 对每个元素做指数运算
        partition = X_exp.sum(dim=1, keepdim=True)  # 求列和，即对同行元素求和 n*1
        return X_exp / partition  # broadcast

    def forward(self, x):
        x_ = x.view((-1, num_inputs))
        y_ = self.linear(x_)
        y_hat = self.softmax(y_)
        return y_hat
# 由评论区指出这里代码有点问题，不需要进行softmax操作，因为损失函数CrossEntropyLoss()默认使用了内置的softmax

四、损失函数与优化器

net = softmaxnet(num_inputs, num_outputs)
lr = 0.1
loss = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=lr)

五、测试集的准确度与损失计算

def get_test_info(data_iter, net):
    right_count, all_count = 0.0, 0
    for x, y in data_iter:
        y_ = net(x)
        l = loss(y_, y)
        right_count += (y_.argmax(dim=1)==y).sum().item()
        all_count += y.shape[0]
    return right_count/all_count, l.item()

六、模型训练及测试

num_epoch = 5


for epoch in range(num_epoch):
    train_r_num, train_all_num = 0.0, 0
    for X, y in tqdm(train_iter):
        y_ = net(X)
        l = loss(y_, y)
        l.backward()
        optimizer.step()
        optimizer.zero_grad()
        train_r_num += (y_.argmax(dim=1) == y).sum().item()
        train_all_num += y.shape[0]
    test_acc, test_ave_loss = get_test_info(test_iter, net)
    print('epoch %d, train loss %.4f, train acc %.3f' % (epoch+1, l.item(), train_r_num/train_all_num))
    print('test loss %.4f, test acc %.3f' % (test_ave_loss, test_acc))

实验结果