[FIXED] PyTorch MaxPool2D unexpected behavior with padding=1

Issue

I was playing around with MaxPool2D in PyTorch and discovered strange behavior when setting padding=1. Here is what I got:

Code:

import torch
from torch.nn.functional import max_pool2d

TEST = 1


def test_maxpool(negative=False, tnsr_size=2, kernel_size=2, stride=2, padding=0):
    """Test MaxPool2D.
    """
    global TEST
    print(f'=== TEST {TEST} ===')
    print(*[f'{i[0]}: {i[1]}' for i in locals().items()], sep=' | ')

    inp = torch.arange(1., tnsr_size ** 2 + 1).reshape(1, tnsr_size, tnsr_size)
    inp = -inp if negative else inp

    print('In:')
    print(inp)

    out = max_pool2d(inp, kernel_size, stride, padding=padding)
    print('Out:')
    print(out)
    print()
    TEST += 1


test_maxpool()
test_maxpool(True)

test_maxpool(padding=1)
test_maxpool(True, padding=1)

Out:

=== TEST 1 ===
negative: False | tnsr_size: 2 | kernel_size: 2 | stride: 2 | padding: 0
In:
tensor([[[1., 2.],
         [3., 4.]]])
Out:
tensor([[[4.]]])

=== TEST 2 ===
negative: True | tnsr_size: 2 | kernel_size: 2 | stride: 2 | padding: 0
In:
tensor([[[-1., -2.],
         [-3., -4.]]])
Out:
tensor([[[-1.]]])

=== TEST 3 ===
negative: False | tnsr_size: 2 | kernel_size: 2 | stride: 2 | padding: 1
In:
tensor([[[1., 2.],
         [3., 4.]]])
Out:
tensor([[[1., 2.],
         [3., 4.]]])

=== TEST 4 ===
negative: True | tnsr_size: 2 | kernel_size: 2 | stride: 2 | padding: 1
In:
tensor([[[-1., -2.],
         [-3., -4.]]])
Out:
tensor([[[-1., -2.],
         [-3., -4.]]])

Tests 1, 2, 3 are fine but Test 4 is weird, I expected to get [[0 0], [0 0]] tensor:

In:
[[-1 -2]
 [-3 -4]]

+ padding -> 

[[ 0  0  0  0]
 [ 0 -1 -2  0]
 [ 0 -3 -4  0]
 [ 0  0  0  0]]

-> kernel_size=2, stride=2 ->

[[0 0]
 [0 0]]

According to Test 3 zero padding was used but Test 4 produced controversial result.

What kind of padding (if any) was that? Why does MaxPool2D behave like that?

pytorch 1.3.1

Solution

This was expected behavior since negative infinity padding is done by default.

The documentation for MaxPool is now fixed. See this PR: Fix MaxPool default pad documentation #59404 .

Answered By - trsvchn