Tensor views

A view is a tensor that shares its underlying data with another tensor. Reshaping, selecting dimensions, or reversing a tensor are all cheap — instead of copying the elements, the view stores different metadata (shape, strides) over the same storage. This is why operations like view, transpose, and squeeze are instant and memory-efficient.

import tensorplay as tp

x = tp.arange(12).reshape(3, 4)
y = x.t()          # a transposed view
print(x.data_ptr() == y.data_ptr())  # True — same underlying data
y[0, 0] = 99
print(x[0, 0].item())   # 99 — the change is visible through the view

The rule of thumb: an operation that only changes shape, strides, or layout produces a view; an operation that creates new elements copies. reshape may return a view or a copy depending on the memory layout — use view when you know the tensor is contiguous and want to guarantee no copy.

x = tp.arange(6).reshape(2, 3)
print(x.transpose(0, 1).shape)   # (3, 2)
print(x.view(6))                 # flatten, no copy
print(x.narrow(1, 0, 2).shape)   # (2, 2) — rows 0:2 of the second dim
b = tp.ones(1, 3)
print(b.expand(4, 3).shape)      # (4, 3) — broadcast a size-1 dim, no copy