data API

MNIST Loader

from MiniTorch.data.mnist import load_mnist

Downloads and caches the MNIST dataset, returning NumPy arrays.

Usage

(x_train, y_train), (x_test, y_test) = load_mnist(normalize=True, flatten=True)

Parameter	Default	Description
normalize	True	Scale pixel values to [0, 1]
flatten	True	Flatten 28×28 images to 784-dim vectors

Returned shapes:

Array	Shape
x_train	(60000, 784)
y_train	(60000,) — integer labels 0–9
x_test	(10000, 784)
y_test	(10000,)

---

DataLoader

from MiniTorch.data.dataloader import DataLoader

Mini-batch iterator with optional shuffling.

Constructor

DataLoader(x, y, batch_size=32, shuffle=True)

Parameter	Default	Description
x	—	Input array (N, ...)
y	—	Label array (N,)
batch_size	32	Number of samples per batch
shuffle	True	Shuffle before each epoch

Iteration

loader = DataLoader(x_train, y_train, batch_size=64, shuffle=True)

for x_batch, y_batch in loader:
    # x_batch: np.ndarray (64, 784)
    # y_batch: np.ndarray (64,)
    ...

Each x_batch / y_batch is a plain NumPy array — wrap in Variable as needed:

for x_np, y_np in loader:
    x = Variable(x_np)
    logits = model(x)
    loss = softmax_cross_entropy(logits, y_np)