注意

单击此处下载完整的示例代码

学习基础知识 || 快速入门 || 张量 || 数据集与数据加载器 || 变换 || 构建模型 || 自动微分 || 优化 || 保存和加载模型

数据集与数据加载器¶

处理数据样本的代码可能很混乱且难以维护；理想情况下，我们希望将数据集代码与模型训练代码分离，以提高可读性和模块化。PyTorch 提供了两个数据基本类型：torch.utils.data.DataLoader 和 torch.utils.data.Dataset，它们使您可以使用预加载的数据集以及您自己的数据。Dataset 存储样本及其对应的标签，DataLoader 在 Dataset 周围包装一个可迭代对象，以方便访问样本。

PyTorch 领域库提供了许多预加载的数据集（例如 FashionMNIST），这些数据集是 torch.utils.data.Dataset 的子类，并实现了特定于该数据的函数。它们可以用于对模型进行原型设计和基准测试。您可以在以下位置找到它们：图像数据集、文本数据集和音频数据集

加载数据集¶

以下是如何从 TorchVision 加载 Fashion-MNIST 数据集的示例。Fashion-MNIST 是一个 Zalando 服装图像数据集，包含 60,000 个训练样本和 10,000 个测试样本。每个样本包含一个 28×28 的灰度图像，以及一个与 10 个类别中的一个相关的标签。

我们使用以下参数加载 FashionMNIST 数据集

root 是存放训练/测试数据的路径。
train 指定训练或测试数据集。
download=True 如果 root 中没有数据，则从互联网下载数据。
transform 和 target_transform 指定特征和标签转换。

import torch
from torch.utils.data import Dataset
from torchvision import datasets
from torchvision.transforms import ToTensor
import matplotlib.pyplot as plt


training_data = datasets.FashionMNIST(
    root="data",
    train=True,
    download=True,
    transform=ToTensor()
)

test_data = datasets.FashionMNIST(
    root="data",
    train=False,
    download=True,
    transform=ToTensor()
)

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz to data/FashionMNIST/raw/train-images-idx3-ubyte.gz

  0%|          | 0.00/26.4M [00:00<?, ?B/s]
  0%|          | 65.5k/26.4M [00:00<01:13, 361kB/s]
  1%|          | 164k/26.4M [00:00<00:56, 466kB/s]
  1%|          | 262k/26.4M [00:00<00:52, 500kB/s]
  1%|1         | 393k/26.4M [00:00<00:44, 586kB/s]
  2%|1         | 524k/26.4M [00:00<00:40, 635kB/s]
  3%|2         | 688k/26.4M [00:01<00:35, 724kB/s]
  3%|3         | 852k/26.4M [00:01<00:32, 782kB/s]
  4%|3         | 1.02M/26.4M [00:01<00:30, 820kB/s]
  5%|4         | 1.21M/26.4M [00:01<00:27, 901kB/s]
  5%|5         | 1.44M/26.4M [00:01<00:24, 1.01MB/s]
  6%|6         | 1.67M/26.4M [00:01<00:20, 1.18MB/s]
  7%|6         | 1.84M/26.4M [00:02<00:20, 1.18MB/s]
  8%|7         | 2.10M/26.4M [00:02<00:19, 1.27MB/s]
  9%|9         | 2.39M/26.4M [00:02<00:15, 1.51MB/s]
 10%|9         | 2.59M/26.4M [00:02<00:16, 1.48MB/s]
 11%|#1        | 2.92M/26.4M [00:02<00:13, 1.75MB/s]
 12%|#1        | 3.15M/26.4M [00:02<00:13, 1.72MB/s]
 13%|#3        | 3.54M/26.4M [00:02<00:11, 2.07MB/s]
 14%|#4        | 3.80M/26.4M [00:03<00:11, 2.01MB/s]
 16%|#5        | 4.23M/26.4M [00:03<00:09, 2.35MB/s]
 17%|#7        | 4.52M/26.4M [00:03<00:09, 2.29MB/s]
 19%|#9        | 5.05M/26.4M [00:03<00:07, 2.76MB/s]
 20%|##        | 5.37M/26.4M [00:03<00:07, 2.64MB/s]
 23%|##2       | 6.00M/26.4M [00:03<00:06, 3.22MB/s]
 24%|##4       | 6.39M/26.4M [00:03<00:06, 3.11MB/s]
 27%|##6       | 7.08M/26.4M [00:04<00:05, 3.71MB/s]
 29%|##8       | 7.54M/26.4M [00:04<00:05, 3.58MB/s]
 32%|###1      | 8.32M/26.4M [00:04<00:04, 4.26MB/s]
 33%|###3      | 8.85M/26.4M [00:04<00:04, 4.11MB/s]
 37%|###7      | 9.80M/26.4M [00:04<00:03, 4.98MB/s]
 39%|###9      | 10.4M/26.4M [00:04<00:03, 4.76MB/s]
 43%|####3     | 11.5M/26.4M [00:04<00:02, 5.74MB/s]
 46%|####6     | 12.2M/26.4M [00:05<00:02, 5.50MB/s]
 51%|#####     | 13.4M/26.4M [00:05<00:01, 6.62MB/s]
 54%|#####3    | 14.2M/26.4M [00:05<00:01, 6.39MB/s]
 59%|#####9    | 15.7M/26.4M [00:05<00:01, 7.68MB/s]
 63%|######2   | 16.6M/26.4M [00:05<00:01, 7.42MB/s]
 69%|######9   | 18.3M/26.4M [00:05<00:00, 8.86MB/s]
 73%|#######3  | 19.4M/26.4M [00:05<00:00, 8.58MB/s]
 80%|########  | 21.3M/26.4M [00:05<00:00, 10.6MB/s]
 85%|########4 | 22.4M/26.4M [00:06<00:00, 10.6MB/s]
 90%|########9 | 23.7M/26.4M [00:06<00:00, 10.6MB/s]
 94%|#########4| 24.9M/26.4M [00:06<00:00, 10.9MB/s]
100%|##########| 26.4M/26.4M [00:06<00:00, 4.14MB/s]
Extracting data/FashionMNIST/raw/train-images-idx3-ubyte.gz to data/FashionMNIST/raw

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz to data/FashionMNIST/raw/train-labels-idx1-ubyte.gz

  0%|          | 0.00/29.5k [00:00<?, ?B/s]
100%|##########| 29.5k/29.5k [00:00<00:00, 327kB/s]
Extracting data/FashionMNIST/raw/train-labels-idx1-ubyte.gz to data/FashionMNIST/raw

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz to data/FashionMNIST/raw/t10k-images-idx3-ubyte.gz

  0%|          | 0.00/4.42M [00:00<?, ?B/s]
  1%|1         | 65.5k/4.42M [00:00<00:12, 359kB/s]
  4%|3         | 164k/4.42M [00:00<00:09, 464kB/s]
  6%|5         | 262k/4.42M [00:00<00:08, 498kB/s]
  9%|8         | 393k/4.42M [00:00<00:06, 583kB/s]
 12%|#1        | 524k/4.42M [00:00<00:06, 632kB/s]
 15%|#4        | 655k/4.42M [00:01<00:05, 661kB/s]
 19%|#8        | 819k/4.42M [00:01<00:04, 737kB/s]
 23%|##2       | 1.02M/4.42M [00:01<00:04, 844kB/s]
 27%|##7       | 1.21M/4.42M [00:01<00:03, 916kB/s]
 32%|###1      | 1.41M/4.42M [00:01<00:03, 966kB/s]
 37%|###7      | 1.64M/4.42M [00:02<00:02, 1.05MB/s]
 43%|####2     | 1.90M/4.42M [00:02<00:02, 1.17MB/s]
 50%|####9     | 2.20M/4.42M [00:02<00:01, 1.30MB/s]
 57%|#####7    | 2.52M/4.42M [00:02<00:01, 1.45MB/s]
 65%|######5   | 2.88M/4.42M [00:02<00:00, 1.60MB/s]
 73%|#######3  | 3.24M/4.42M [00:02<00:00, 1.72MB/s]
 83%|########2 | 3.67M/4.42M [00:03<00:00, 1.90MB/s]
 94%|#########4| 4.16M/4.42M [00:03<00:00, 2.13MB/s]
100%|##########| 4.42M/4.42M [00:03<00:00, 1.31MB/s]
Extracting data/FashionMNIST/raw/t10k-images-idx3-ubyte.gz to data/FashionMNIST/raw

Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz
Downloading http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz to data/FashionMNIST/raw/t10k-labels-idx1-ubyte.gz

  0%|          | 0.00/5.15k [00:00<?, ?B/s]
100%|##########| 5.15k/5.15k [00:00<00:00, 31.2MB/s]
Extracting data/FashionMNIST/raw/t10k-labels-idx1-ubyte.gz to data/FashionMNIST/raw

迭代和可视化数据集¶

我们可以像列表一样手动索引 Datasets：training_data[index]。我们使用 matplotlib 来可视化训练数据中的部分样本。

labels_map = {
    0: "T-Shirt",
    1: "Trouser",
    2: "Pullover",
    3: "Dress",
    4: "Coat",
    5: "Sandal",
    6: "Shirt",
    7: "Sneaker",
    8: "Bag",
    9: "Ankle Boot",
}
figure = plt.figure(figsize=(8, 8))
cols, rows = 3, 3
for i in range(1, cols * rows + 1):
    sample_idx = torch.randint(len(training_data), size=(1,)).item()
    img, label = training_data[sample_idx]
    figure.add_subplot(rows, cols, i)
    plt.title(labels_map[label])
    plt.axis("off")
    plt.imshow(img.squeeze(), cmap="gray")
plt.show()

Ankle Boot, Shirt, Bag, Ankle Boot, Trouser, Sandal, Coat, Sandal, Pullover

为您的文件创建自定义数据集¶

自定义 Dataset 类必须实现三个函数：__init__、__len__ 和 __getitem__。请看此实现；FashionMNIST 图像存储在目录 img_dir 中，它们的标签单独存储在 CSV 文件 annotations_file 中。

在接下来的部分中，我们将详细介绍这些函数中的每个函数。

import os
import pandas as pd
from torchvision.io import read_image

class CustomImageDataset(Dataset):
    def __init__(self, annotations_file, img_dir, transform=None, target_transform=None):
        self.img_labels = pd.read_csv(annotations_file)
        self.img_dir = img_dir
        self.transform = transform
        self.target_transform = target_transform

    def __len__(self):
        return len(self.img_labels)

    def __getitem__(self, idx):
        img_path = os.path.join(self.img_dir, self.img_labels.iloc[idx, 0])
        image = read_image(img_path)
        label = self.img_labels.iloc[idx, 1]
        if self.transform:
            image = self.transform(image)
        if self.target_transform:
            label = self.target_transform(label)
        return image, label

`init`¶

__init__ 函数在实例化 Dataset 对象时运行一次。我们初始化包含图像的目录、注释文件和两种转换（将在下一节中详细介绍）。

labels.csv 文件的格式如下

tshirt1.jpg, 0
tshirt2.jpg, 0
......
ankleboot999.jpg, 9

def __init__(self, annotations_file, img_dir, transform=None, target_transform=None):
    self.img_labels = pd.read_csv(annotations_file)
    self.img_dir = img_dir
    self.transform = transform
    self.target_transform = target_transform

`len`¶

__len__ 函数返回数据集中的样本数量。

示例

def __len__(self):
    return len(self.img_labels)

`getitem`¶

__getitem__ 函数加载并返回给定索引 idx 处的数据集中的样本。根据索引，它识别图像在磁盘上的位置，使用 read_image 将其转换为张量，从 self.img_labels 中的 csv 数据中检索相应的标签，对它们调用转换函数（如果适用），并将张量图像和相应的标签以元组形式返回。

def __getitem__(self, idx):
    img_path = os.path.join(self.img_dir, self.img_labels.iloc[idx, 0])
    image = read_image(img_path)
    label = self.img_labels.iloc[idx, 1]
    if self.transform:
        image = self.transform(image)
    if self.target_transform:
        label = self.target_transform(label)
    return image, label

使用 DataLoaders 准备您的训练数据¶

Dataset 一次获取数据集的特征和标签。在训练模型时，我们通常希望以“小批量”的形式传递样本，在每个 epoch 中重新洗牌数据以减少模型过拟合，并使用 Python 的 multiprocessing 来加速数据检索。

DataLoader 是一个可迭代对象，它通过一个简单的 API 为我们抽象了这种复杂性。

from torch.utils.data import DataLoader

train_dataloader = DataLoader(training_data, batch_size=64, shuffle=True)
test_dataloader = DataLoader(test_data, batch_size=64, shuffle=True)

遍历 DataLoader¶

我们已将该数据集加载到 DataLoader 中，并且可以根据需要遍历该数据集。下面的每次迭代都会返回一批 train_features 和 train_labels（分别包含 batch_size=64 个特征和标签）。因为我们指定了 shuffle=True，所以在我们遍历完所有批次后，数据会进行洗牌（要对数据加载顺序进行更细粒度的控制，请查看采样器）。

# Display image and label.
train_features, train_labels = next(iter(train_dataloader))
print(f"Feature batch shape: {train_features.size()}")
print(f"Labels batch shape: {train_labels.size()}")
img = train_features[0].squeeze()
label = train_labels[0]
plt.imshow(img, cmap="gray")
plt.show()
print(f"Label: {label}")

Feature batch shape: torch.Size([64, 1, 28, 28])
Labels batch shape: torch.Size([64])
Label: 5

进一步阅读¶

torch.utils.data API

脚本的总运行时间：（0 分钟 12.729 秒）

由 Sphinx-Gallery 生成的图库

数据集与数据加载器¶

加载数据集¶

迭代和可视化数据集¶

为您的文件创建自定义数据集¶

`init`¶

`len`¶

`getitem`¶

使用 DataLoaders 准备您的训练数据¶

遍历 DataLoader¶

进一步阅读¶

文档

教程

资源

数据集与数据加载器¶

加载数据集¶

迭代和可视化数据集¶

为您的文件创建自定义数据集¶

__init__¶

__len__¶

__getitem__¶

使用 DataLoaders 准备您的训练数据¶

遍历 DataLoader¶

进一步阅读¶

文档

教程

资源

`init`¶

`len`¶

`getitem`¶