正常情况下，torch读取数据的时候是Batch Size小批量读取。
首先找到所有数据集的路径保持到一个变量中，之后需要读取哪个数据的时候，就根据这个变量中的路径索引去读取。因为硬件的限制，从硬盘中读取数据到显存中所花的时间要远远大于从内存中读取数据到显存中。因此，如果程序直接是从硬盘上读取数据到显存中，就会非常耗时。
我们可以把对应的图像直接全部都读取到内存中，然后将其转换为CPU类型的Tensor Dataloader，之后迭代的时候，从这个Tensor中迭代，那么网络在读取数据的时候，就是直接从内存中进行读取了，那么提高的速度就不是一点半点儿了，MNIST训练一个epoch的时间可以由原来的50s减少到5s，速度可以说是提高了10倍。

核心代码

## train_set = torchvision.datasets.MNIST(DOWNLOAD_PATH, train=True, download=True,
                                       transform=transform_mnist)       #train_set其实只是读取的一个路径，但是并没有读取数据主体

train_data = train_set.data.float().unsqueeze(1) / 255.0        #读取所有的数据集路径，然后将对应的数据集读取到内存中
train_label = train_set.targets

train_dataset =  TensorDataset(train_data,train_label)      #将原本保存在内存中的变量转化为Tensor
train_loader = DataLoader(train_dataset,batch_size=batch_size,shuffle=True)

完整代码

代码

### import time

import torch
import torch.nn.functional as F
import torchvision
from einops import rearrange
from torch import nn
from torch import optim
from torch.utils.data import TensorDataset, DataLoader


#残差模块，放在每个前馈网络和注意力之后
class Residual(nn.Module):
    def __init__(self, fn):
        super().__init__()
        self.fn = fn

    def forward(self, x, **kwargs):
        return self.fn(x, **kwargs) + x

#layernorm归一化,放在多头注意力层和激活函数层。用绝对位置编码的BERT，layernorm用来自身通道归一化
class PreNorm(nn.Module):
    def __init__(self, dim, fn):
        super().__init__()
        self.norm = nn.LayerNorm(dim)
        self.fn = fn

    def forward(self, x, **kwargs):
        return self.fn(self.norm(x), **kwargs)
#放置多头注意力后，因为在于多头注意力使用的矩阵乘法为线性变换，后面跟上由全连接网络构成的FeedForward增加非线性结构
class FeedForward(nn.Module):
    def __init__(self, dim, hidden_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dim, hidden_dim),
            nn.GELU(),
            nn.Linear(hidden_dim, dim)
        )

    def forward(self, x):
        return self.net(x)
#多头注意力层，多个自注意力连起来。使用qkv计算
class Attention(nn.Module):
    def __init__(self, dim, heads=8):
        super().__init__()
        self.heads = heads
        self.scale = dim ** -0.5

        self.to_qkv = nn.Linear(dim, dim * 3, bias=False)
        self.to_out = nn.Linear(dim, dim)

    def forward(self, x, mask = None):
        b, n, _, h = *x.shape, self.heads
        qkv = self.to_qkv(x)
        q, k, v = rearrange(qkv, 'b n (qkv h d) -> qkv b h n d', qkv=3, h=h)

        dots = torch.einsum('bhid,bhjd->bhij', q, k) * self.scale

        if mask is not None:
            mask = F.pad(mask.flatten(1), (1, 0), value = True)
            assert mask.shape[-1] == dots.shape[-1], 'mask has incorrect dimensions'
            mask = mask[:, None, :] * mask[:, :, None]
            dots.masked_fill_(~mask, float('-inf'))
            del mask

        attn = dots.softmax(dim=-1)

        out = torch.einsum('bhij,bhjd->bhid', attn, v)
        out = rearrange(out, 'b h n d -> b n (h d)')
        out =  self.to_out(out)
        return out

class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, mlp_dim):
        super().__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                Residual(PreNorm(dim, Attention(dim, heads = heads))),
                Residual(PreNorm(dim, FeedForward(dim, mlp_dim)))
            ]))

    def forward(self, x, mask=None):
        for attn, ff in self.layers:
            x = attn(x, mask=mask)
            x = ff(x)
        return x
#将图像切割成一个个图像块,组成序列化的数据输入Transformer执行图像分类任务。
class ViT(nn.Module):
    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, channels=3):
        super().__init__()
        assert image_size % patch_size == 0, 'image dimensions must be divisible by the patch size'
        num_patches = (image_size // patch_size) ** 2
        patch_dim = channels * patch_size ** 2

        self.patch_size = patch_size

        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
        self.patch_to_embedding = nn.Linear(patch_dim, dim)
        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
        self.transformer = Transformer(dim, depth, heads, mlp_dim)

        self.to_cls_token = nn.Identity()

        self.mlp_head = nn.Sequential(
            nn.Linear(dim, mlp_dim),
            nn.GELU(),
            nn.Linear(mlp_dim, num_classes)
        )

    def forward(self, img, mask=None):
        p = self.patch_size

        x = rearrange(img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = p, p2 = p)
        x = self.patch_to_embedding(x)

        cls_tokens = self.cls_token.expand(img.shape[0], -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)
        x += self.pos_embedding
        x = self.transformer(x, mask)

        x = self.to_cls_token(x[:, 0])
        return self.mlp_head(x)

def train_epoch(model, optimizer, data_loader, loss_history):
    total_samples = len(data_loader.dataset)
    model = model.cuda()
    model.train()

    for i, (data, target) in enumerate(data_loader):
        # print("data.shape:", data.shape)
        data=data.cuda()
        target=target.cuda()
        optimizer.zero_grad()
        output = F.log_softmax(model(data), dim=1)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()

        if i % 100 == 0:
            print('[' +  '{:5}'.format(i * len(data)) + '/' + '{:5}'.format(total_samples) +
                  ' (' + '{:3.0f}'.format(100 * i / len(data_loader)) + '%)]  Loss: ' +
                  '{:6.4f}'.format(loss.item()))
            loss_history.append(loss.item())


def evaluate(model, data_loader, loss_history):
    model.eval()

    total_samples = len(data_loader.dataset)
    correct_samples = 0
    total_loss = 0
    model = model.cuda()
    with torch.no_grad():
        for data, target in data_loader:
            data = data.cuda()
            target = target.cuda()
            output = F.log_softmax(model(data), dim=1)
            loss = F.nll_loss(output, target, reduction='sum')
            _, pred = torch.max(output, dim=1)

            total_loss += loss.item()
            correct_samples += pred.eq(target).sum()

    avg_loss = total_loss / total_samples
    loss_history.append(avg_loss)
    print('\nAverage test loss: ' + '{:.4f}'.format(avg_loss) +
          '  Accuracy:' + '{:5}'.format(correct_samples) + '/' +
          '{:5}'.format(total_samples) + ' (' +
          '{:4.2f}'.format(100.0 * correct_samples / total_samples) + '%)\n')





# 设置随机种子
torch.manual_seed(42)

# 定义MNIST数据集的下载路径
DOWNLOAD_PATH = './data/mnist'

# 定义每个batch中包含的样本数量
batch_size = 5000

# 定义数据集预处理操作
transform_mnist = torchvision.transforms.Compose([
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize((0.1307,), (0.3081,))
])

# 设置数据加载器的线程数
num_workers = 0

# 读入训练集和测试集数据
train_set = torchvision.datasets.MNIST(DOWNLOAD_PATH, train=True, download=True, transform=transform_mnist)
test_set = torchvision.datasets.MNIST(DOWNLOAD_PATH, train=False, download=True, transform=transform_mnist)

# 将数据集转换为张量
train_data = train_set.data.float().unsqueeze(1) / 255.0
train_labels = train_set.targets
test_data = test_set.data.float().unsqueeze(1) / 255.0
test_labels = test_set.targets

# 将数据集包装成Dataset对象
train_dataset = TensorDataset(train_data, train_labels)
test_dataset = TensorDataset(test_data, test_labels)

# 创建训练集和测试集的DataLoader
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)





N_EPOCHS = 100

'''
patch大小为 7x7（对于 28x28 图像，这意味着每个图像 4 x 4 = 16 个patch）、10 个可能的目标类别（0 到 9）和 1 个颜色通道（因为图像是灰度）。
在网络参数方面，使用了 64 个单元的维度，6 个 Transformer 块的深度，8 个 Transformer 头，MLP 使用 128 维度。'''
model = ViT(image_size=28, patch_size=7, num_classes=10, channels=1,
            dim=64, depth=6, heads=8, mlp_dim=128)
optimizer = optim.Adam(model.parameters(), lr=0.003)

train_loss_history, test_loss_history = [], []
for epoch in range(1, N_EPOCHS + 1):
    start_time = time.time()
    print('Epoch:', epoch)
    train_epoch(model, optimizer, train_loader, train_loss_history)
    evaluate(model, test_loader, test_loss_history)

    print('Execution time:', '{:5.2f}'.format(time.time() - start_time), 'seconds')