DataLoader turns a Dataset into an iteration stream for training or testing. Dataset defines how to retrieve one example; DataLoader chooses examples, groups them into batches, and can fetch upcoming data while the model processes the current batch. section 2.1.4 section 2.7
Configure the loader according to the execution workload:
- Set shuffle=True for training so the model does not learn the dataset's original ordering; omit it for deterministic test evaluation. section 2.1.4 section 3.6
- Use batch_size to trade larger computational batches against the number of optimization steps. section 2.7
- Use num_workers to preload batches, especially when CPU transforms would otherwise leave the GPU waiting. section 3.6
- Prefer iteration over keeping a complete dataset in memory when data may exceed available memory. section 2.7
Minimal regression loader
dataset = Simple1DRegressionDataset(X, y) training_loader = DataLoader(dataset, shuffle=True) The custom Dataset reshapes inputs and labels, returns float32 tensors, and supplies individual indexed examples; the DataLoader randomizes their retrieval for training. section 2.1.4
train_transform = transforms.Compose([
transforms.RandomAffine(degrees=5, translate=(0.05, 0.05),
scale=(0.98, 1.02)),
transforms.ToTensor(),
])
test_transform = transforms.ToTensor()
mnist_train_t = torchvision.datasets.MNIST("./data", train=True,
transform=train_transform)
mnist_test_t = torchvision.datasets.MNIST("./data", train=False,
transform=test_transform)
mnist_train_loader_t = DataLoader(mnist_train_t, shuffle=True,
batch_size=B, num_workers=5)
mnist_test_loader_t = DataLoader(mnist_test_t, batch_size=B,
num_workers=5)
Inspect a training/test setup that combines on-the-fly image augmentation, batching, shuffling, deterministic testing, and worker-based preloading. section 3.6
