Function Differences with torch.nn.AvgPool2d

torch.nn.AvgPool2d

torch.nn.AvgPool2d(kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None)(input) -> Tensor

For more information, see torch.nn.AvgPool2d.

mindspore.nn.AvgPool2d

mindspore.nn.AvgPool2d(kernel_size=1, stride=1, pad_mode='valid', padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None, data_format='NCHW')(x) -> Tensor

For more information, see mindspore.nn.AvgPool2d.

Differences

PyTorch: Apply two-dimensional averaging pooling to an input signal consisting of multiple input planes.

MindSpore: This API implementation function of MindSpore is compatible with TensorFlow and PyTorch, When pad_mode is “valid” or “same”, the function is consistent with TensorFlow, and when pad_mode is “pad”, the function is consistent with PyTorch, compared with PyTorch 1.8.1, MindSpore additionally supports 3D input, which is consistent with PyTorch 1.12.

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	kernel_size	kernel_size	Consistent function, no default values for PyTorch
	Parameter 2	stride	stride	Consistent function, different default values of parameters
	Parameter 3	padding	padding	Consistent
	Parameter 4	ceil_mode	ceil_mode	Consistent
	Parameter 5	count_include_pad	count_include_pad	Consistent
	Parameter 6	divisor_override	divisor_override	Consistent
	Parameter 7	-	pad_mode	MindSpore specifies how the pooling will be filled, with optional values of “same”, “valid” or “pad”. PyTorch does not have this parameter
	Parameter 8	-	data_format	Specify the input data format in MindSpore, either “NHWC” or “NCHW”. PyTorch does not have this parameter
Input	Single input	input	x	Same function, different parameter names

Code Example 1

The two APIs achieve the same function and have the same usage.

# PyTorch
import torch
import torch.nn as nn

m = nn.AvgPool2d(kernel_size=1, stride=1)
input_x = torch.tensor([[[[1, 0, 1], [0, 1, 1]]]],dtype=torch.float32)
output = m(input_x)
print(output.numpy())
# [[[[1. 0. 1.]
#    [0. 1. 1.]]]]

# MindSpore
import mindspore
import mindspore.nn as nn
from mindspore import Tensor

pool = nn.AvgPool2d(kernel_size=1, stride=1)
x = Tensor([[[[1, 0, 1], [0, 1, 1]]]], dtype=mindspore.float32)
output = pool(x)
print(output)
# [[[[1. 0. 1.]
#    [0. 1. 1.]]]]

Code Example 2

Use pad mode to ensure functional consistency.

import torch
import mindspore.nn as nn
import mindspore.ops as ops

pool = nn.AvgPool2d(4, stride=1, ceil_mode=True, pad_mode='pad', padding=2)
x1 = ops.randn(6, 6, 8, 8)
output = pool(x1)
print(output.shape)
# (6, 6, 9, 9)

pool = torch.nn.AvgPool2d(4, stride=1, ceil_mode=True, padding=2)
x1 = torch.randn(6, 6, 8, 8)
output = pool(x1)
print(output.shape)
# torch.Size([6, 6, 9, 9])