[pytorch learning notes] datasets and dataloaders

Why set it separately Dataloaders

pytorch We hope to separate the data set code from the model training code , In order to get better readability and modularity .pytorch Provide two interface functions :torch.utils.data.Dataloader and torch.utils.data.Dataset. Load your own dataset and pytorch Built in datasets .

Load data set

Here is a demonstration of how to torchvision load fashion-mnist An example of a data set , The code is as follows ：

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', # root It's storage training / The path of the test data 
    train = True, #  Specify training or test 
    download=True, #  If data is not available , Then download data from the Internet to root
    transform=ToTensor()
)

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

Iterative and visual data sets

We can get Datasets Index manually like a list , The visualization code is as follows ：

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()

Create a custom dataset

Customize dataset Class must implement three functions ,_init、len__ and __getitem. take fashionmnist Images are stored in a directory img_dir, Their labels are stored in a CSV In file . The code is as follows ：

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):
    #  initialization 
        self.img_labels = pd.read_csv(annotations_file) # Read picture labels 
        self.img_dir = img_dir  #  Picture path 
        self.transform = transform  #  Picture enhancement field 
        self.target_transform = target_transform #  Label enhanced fields 

    def __len__(self):
    # Returns the number of samples in the dataset 
        return len(self.img_labels) #

    def __getitem__(self, idx):
    #  Read a single sample and call the image transformation function , Return the transformed function 
        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

Use DataLoaders Prepare data for training

In training the model , Usually mini-batch To train , At every epoch Reshuffle to reduce model overfitting , And use python multiprocessing Accelerate data retrieval .dataloader Is an iterative object ：

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)

Traverse DataLoader

Data set loaded into DataLoader You can traverse the data set as needed , Each iteration below will return 64 Images and labels .

# Display image and label.
train_features, train_labels = next(iter(train_dataloader)) # 64 Image and label data 
print(f"Feature batch shape: {
      train_features.size()}")
print(f"Labels batch shape: {
      train_labels.size()}")
img = train_features[0].squeeze()#  Select the first image , And will batch Dimension compression 
label = train_labels[0]
plt.imshow(img, cmap="gray")
plt.show()
print(f"Label: {
      label}")