Reshape

reshape() and view() change the tensor's shape.

Reshape and view

reshape and view change the shape while keeping the number of elements; -1 lets a dimension be worked out automatically. view only works on tensors that are contiguous in memory and always shares the data; reshape copies if needed:

x = torch.arange(12)
print(x.view(3, 4).shape)        # torch.Size([3, 4])
print(x.reshape(2, -1).shape)    # torch.Size([2, 6])
img = torch.zeros(32, 1, 28, 28)
print(img.view(32, -1).shape)    # torch.Size([32, 784]): flattening

When in doubt, use reshape; it works in every case.

Transpose and permute

transpose(a, b) swaps two axes, permute reorders all axes. Images are in (height, width, channel) order in NumPy/PIL and (channel, height, width) in PyTorch:

hwc = torch.zeros(28, 28, 3)
chw = hwc.permute(2, 0, 1)
print(chw.shape)               # torch.Size([3, 28, 28])
print(chw.is_contiguous())     # False: reshape may be needed instead of view

After permute the tensor isn't contiguous; it is fixed with .contiguous().

Exercise Image shapes

Convert an HWC image batch of shape (4, 28, 28, 3) to NCHW with permute(0, 3, 1, 2) and print its shape. Then flatten it with reshape(4, -1) and print its shape, and finally print that it equals the shape of x.flatten(1).

Show example solution

import torch

nhwc = torch.zeros(4, 28, 28, 3)

x = nhwc.permute(0, 3, 1, 2)
print(x.shape)
d = x.reshape(4, -1)
print(d.shape)
print(d.shape == x.flatten(1).shape)

Test yourself

What shape does <code>x.view(8, -1)</code> give for a tensor of shape (8, 3, 4, 4)?
- (8, 48)
- (8, 12)
- (8, 3)
- (384,)
Show answer

(8, 48) — 3 × 4 × 4 = 48; the batch dimension is kept.

Convert the image from HWC to CHW.

chw = hwc.____(2, 0, 1)

Show answer

chw = hwc.permute(2, 0, 1)

Match the operations.

view

Shared reshape
reshape

Copies if needed
permute

Axis order
flatten(1)

Flattens all but batch

Next lesson: Indexing

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies