Tensor

The core Tensor class wraps a NumPy (or CuPy) array and tracks gradients for backpropagation.

Constructor

Tensor(data, requires_grad=False)

Properties

• shape, ndim, dtype - standard array properties
• T - transpose (swaps last two dims)
• device - "cpu" or "cuda"
• size(dim=None) - total elements or size of a specific dimension

Static Constructors

Tensor.zeros(2, 3)
Tensor.ones(2, 3, requires_grad=True)
Tensor.randn(4, 5, device="cuda")
Tensor.eye(3)

Arithmetic

+, -, *, /, ** with broadcasting. Right-hand variants (5.0 + tensor) supported. @ for matrix multiplication.

Reductions

• sum(axis, keepdims), mean(axis, keepdims)
• max(axis, keepdims), min(axis, keepdims)
• var(axis, keepdims), std(axis, keepdims)

Shape Operations

• reshape(*shape), transpose(dim0, dim1)
• squeeze(axis), unsqueeze(axis)
• __getitem__ for indexing/slicing

Elementwise

• exp(), log(), abs(), clamp(min_val, max_val)

Autograd

• backward() - reverse-mode autodiff (scalar tensors only)
• zero_grad(), detach(), clone()

Device Transfer

x = Tensor.randn(3, 4)
x_gpu = x.to("cuda")

no_grad Context Manager

Disable gradient tracking for inference:

from minitorch import no_grad

with no_grad():
    y = model(x)  # no gradients tracked

Free Functions

from minitorch import cat, stack

c = cat([a, b], axis=0)
s = stack([a, b], axis=0)