DataLoader wraps a dataset or iterable and exposes its samples as batches for training. Set batch_size to control batch size and shuffle=True to randomize observations before batching, reducing ordering correlations and stabilizing optimization. section 2.3.1 section 3.2.2
Instead of retrieving individual samples directly, iterate over a DataLoader to feed consistently sized groups to a model. It supports separate training, validation, and test iterators, while batch size must fit the data and available memory. section 2.4.1 section 6.2.2
Batching a binary clothing dataset
Filter Fashion MNIST to t-shirts and ankle boots, then create training and test loaders with batches of 64. Training observations are shuffled so labels are more evenly distributed across iterations. section 2.3.1
Listing 2.4 Creating batches for training and testing
batch_size=64
binary_train_loader=torch.utils.data.DataLoader(
binary_train_set, #1
batch_size=batch_size, #2
shuffle=True) #3
binary_test_loader=torch.utils.data.DataLoader(
binary_test_set, #4
batch_size=batch_size,shuffle=True)
The listing shows the minimal PyTorch pattern for constructing separate training and test DataLoader instances. section 2.3.1
