Why nn.MaxPool2d cannot work on rank 2 tensor?

import torch
import torch.nn as nn
import torch.nn.functional as F

# input = torch.LongTensor(4,4).random_(0, 50)
input = torch.randn(4,4)


m = nn.MaxPool2d(kernel_size=2, stride=2)
output = m(input)

I created the example that will not work, but when I set input = torch.randn(1,4,4) it will work. Can someone explain the logic behind this decision?

I know that back few days ago, when I loaded the image it typically had (1,128,128) dimension if it was single channel or (3,128,128) if it had 3 channels (RGB).

Is this perhaps the logic why we insist that MaxPool2d requires rank 3 or 4 tensors?


Yes, MaxPool2d is built to work with images which are 3D Tensors for a single image and 4D tensors for a batch of images.

1 Like