Name: MaxPool3d() in PyTorch
Rating: 3.9 (602 reviews)
Author: hyperkai

*Memos:

My post explains Pooling Layer.
My post explains MaxPool1d().
My post explains MaxPool2d().
My post explains AvgPool1d().
My post explains AvgPool2d().
My post explains AvgPool3d().
My post explains requires_grad.

MaxPool3d() can get the 4D or 5D tensor of the one or more elements computed by 3D max pooling from the 4D or 5D tensor of one or more elements as shown below:

*Memos:

The 1st argument for initialization is kernel_size(Required-Type:int or tuple or list of int). *It must be 1 <= x.
The 2nd argument for initialization is stride(Optional-Default:None-Type:int or tuple or list of int): *Memos:
- It must be 1 <= x.
- If it's None, kernel_size is set.
The 3rd argument for initialization is padding(Optional-Default:0-Type:int or tuple or list of int). *It must be 0 <= x.
The 4th argument for initialization is dilation(Optional-Default:1-Type:int or tuple or list of int). *It must be 1 <= x.
The 5th argument for initialization is return_indices(Optional-Default:False-Type:bool).
The 6th argument for initialization is ceil_mode(Optional-Default:False-Type:bool).
The 1st argument is input(Required-Type:tensor of float).
The tensor's requires_grad which is False by default is not set to True by MaxPool3d().

import torch
from torch import nn

tensor1 = torch.tensor([[[[8., -3., 0., 1., 5., -2.]]]])

tensor1.requires_grad
# False

maxpool3d = nn.MaxPool3d(kernel_size=1)
tensor2 = maxpool3d(input=tensor1)
tensor2
# tensor([[[[8., -3., 0., 1., 5., -2.]]]])

tensor2.requires_grad
# False

maxpool3d
# MaxPool3d(kernel_size=1, stride=1, padding=0, dilation=1, ceil_mode=False)

maxpool3d.kernel_size
# 1

maxpool3d.stride
# 1

maxpool3d.padding
# 0

maxpool3d.dilation
# 1

maxpool3d.return_indices
# False

maxpool3d.ceil_mode
# False

maxpool3d = nn.MaxPool3d(kernel_size=1, stride=None, padding=0, 
                         dilation=1, return_indices=False, ceil_mode=False)
maxpool3d(input=tensor1)
# tensor([[[[8., -3., 0., 1., 5., -2.]]]])

maxpool3d = nn.MaxPool3d(kernel_size=2, padding=1, return_indices=True)
maxpool3d(input=tensor1)
# (tensor([[[[8., 0., 5., -2.]]]]), tensor([[[[0, 2, 4, 5]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=3, padding=1, return_indices=True)
maxpool3d(input=tensor1)
# (tensor([[[[8., 5.]]]]), tensor([[[[0, 4]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=4, padding=2, return_indices=True)
maxpool3d(input=tensor1)
# (tensor([[[[8., 5.]]]]), tensor([[[[0, 4]]]]))
etc.

maxpool3d = nn.MaxPool3d(kernel_size=7, padding=3, return_indices=True)
maxpool3d(input=tensor1)
# (tensor([[[[8.]]]]), tensor([[[[0]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=8, padding=4, return_indices=True)
maxpool3d(input=tensor1)
# (tensor([[[[8.]]]]), tensor([[[[0]]]]))
etc.

my_tensor = torch.tensor([[[[8., -3., 0.],
                            [1., 5., -2.]]]])
maxpool3d = nn.MaxPool3d(kernel_size=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8., -3., 0.],
#            [1., 5., -2.]]]]),
#  tensor([[[[0, 1, 2],
#            [3, 4, 5]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=2, padding=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8., 0.],
#            [1., 5.]]]]),
#  tensor([[[[0, 2],
#            [3, 4]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=3, padding=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.]]]]),
#  tensor([[[[0]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=4, padding=2, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.]]]]),
#  tensor([[[[0]]]]))
etc.

my_tensor = torch.tensor([[[[8.], [-3.], [0.], [1.], [5.], [-2.]]]])

maxpool3d = nn.MaxPool3d(kernel_size=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.], [-3.], [0.], [1.], [5.], [-2.]]]]),
#  tensor([[[[0], [1], [2], [3], [4], [5]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=2, padding=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.], [0.], [5.], [-2.]]]]),
#  tensor([[[[0], [2], [4], [5]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=3, padding=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.], [5.]]]]),
#  tensor([[[[0], [4]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=4, padding=2, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.], [5.]]]]),
#  tensor([[[[0], [4]]]]))
etc.

maxpool3d = nn.MaxPool3d(kernel_size=7, padding=3, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.]]]]), tensor([[[[0]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=8, padding=4, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[8.]]]]), tensor([[[[0]]]]))
etc.

my_tensor = torch.tensor([[[[[8.], [-3.], [0.]],
                            [[1.], [5.], [-2.]]]]])
maxpool3d = nn.MaxPool3d(kernel_size=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[[8.], [-3.], [0.]],
#            [[1.], [5.], [-2.]]]]]),
#  tensor([[[[[0], [1], [2]],
#            [[3], [4], [5]]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=2, padding=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[[8.], [0.]], [[1.], [5.]]]]]),
#  tensor([[[[[0], [2]], [[3], [4]]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=3, padding=1, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[[8.]]]]]), tensor([[[[[0]]]]]))

maxpool3d = nn.MaxPool3d(kernel_size=4, padding=2, return_indices=True)
maxpool3d(input=my_tensor)
# (tensor([[[[[8.]]]]]), tensor([[[[[0]]]]]))
etc.