layers.FlattenLayer(),
nn.Linear(512, 10))
# 下面用分步add的方法构造，看起来比较丑但是输出的时候名称清晰
net = nn.Sequential(nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3),
nn.BatchNorm2d(64), nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
net.add_module("resnet_block1", resnet_block(64, 64, 2, first_block=True))
net.add_module("resnet_block2", resnet_block(64, 128, 2))
net.add_module("resnet_block3", resnet_block(128, 256, 2))
net.add_module("resnet_block4", resnet_block(256, 512, 2))
net.add_module("global_avg_pool", layers.GlobalAvgPool2d())
net.add_module("fc",
nn.Sequential(layers.FlattenLayer(), nn.Linear(512, 10)))
X = torch.rand(1, 1, 224, 224)
for name, layer in net.named_children():
X = layer(X)
print(name, ' output shape: ', X.shape)
print('————————————————————————————')
# 最后类似之前，进行一下测试，这里也减小了图片大小
# 1epoch = 189.7sec
batch_size = 256
train_iter, test_iter = data_process.load_data_fashion_mnist(batch_size,
resize=96)
lr, num_epochs = 0.001, 5
optim = torch.optim.Adam(net.parameters(), lr=lr)
train.train_ch5(net, train_iter, test_iter, batch_size, optim, device,
num_epochs)
# 同上操作的测试集，注意train=False
mnist_test = torchvision.datasets.FashionMNIST(
root=r"./Datasets", train=False,
download=True, transform=transforms.ToTensor())
# 可以用len得到数据集的大小，用type得到数据集的类型，用下标得到某个具体的数据
print(len(mnist_train), len(mnist_test))
feature, label = mnist_train[10]
# 能看到图片为1*28*28，是一维的灰度图
print(feature.shape, label)
# 编写好用索引获取标签真名的函数get_fashion_mnist_labels
# 和绘制数据图片的show_fashion_mnist后，下面来尝试显示一下
# 类似之前，这里也有特征X和标签y
X, y = [], []
for i in range(10):
# 这个数据集的数据组织方式就是第0是特征数据，第1是标签
X.append(mnist_train[i][0])
y.append(mnist_train[i][1])
# 将这个得到的特征和标签转换为图片绘制
plot.show_fashion_mnist(X, data_process.get_fashion_mnist_labels(y))
# 将读取数据的流程封装到load_data_fashion_mnist中使用
train_iter, _ = data_process.load_data_fashion_mnist(256)
# 测试下读取完数据所需的时间
start = time.time()
for X, y in train_iter:
continue
print('%.2f s' % (time.time()-start))