Differences with torch.nn.MaxPool2d

View Source On Gitee

torch.nn.MaxPool2d

torch.nn.MaxPool2d(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)(input) -> Tensor

For more information, see torch.nn.MaxPool2d.

mindspore.nn.MaxPool2d

mindspore.nn.MaxPool2d(kernel_size=1, stride=1, pad_mode="valid", padding=0, dilation=1, return_indices=False, ceil_mode=False, data_format="NCHW")(x) -> Tensor

For more information, see mindspore.nn.MaxPool2d.

Differences

PyTorch: Perform two-dimensional maximum pooling operations on the input multidimensional data.

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, MindSpore additionally supports 2D input, which is consistent with PyTorch 1.12.

Categories

Subcategories

PyTorch

MindSpore

Difference

Parameters

Parameter 1

kernel_size

kernel_size

Consistent function, no default values for PyTorch

Parameter 2

stride

stride

Consistent function, different default value

Parameter 3

padding

padding

Consistent

Parameter 4

dilation

dilation

Consistent

Parameter 5

return_indices

return_indices

Consistent

Parameter 6

ceil_mode

ceil_mode

Consistent

Parameter 7

input

x

Consistent function, different parameter names

Parameter 8

-

pad_mode

Control the padding mode, and PyTorch does not have this parameter

Parameter 9

-

data_format

The input data format can be "NHWC" or "NCHW". Default value: "NCHW"

Code Example 1

Construct a pooling layer with a convolution kernel size of 1x3 and a step size of 1. The padding defaults to 0 and no element filling is performed. The default value of dilation is 1, and the elements in the window are contiguous. Pooling padding mode returns the output from a valid calculation without padding, and excess pixels that do not satisfy the calculation are discarded. With the same parameter settings, the two APIs achieve the same function to perform two-dimensional maximum pooling operations on the input multidimensional data.

# PyTorch
import torch
from torch import tensor
import numpy as np

pool = torch.nn.MaxPool2d(kernel_size=3, stride=1)
x = tensor(np.random.randint(0, 10, [1, 2, 4, 4]), dtype=torch.float32)
output = pool(x)
result = output.shape
print(tuple(result))
# (1, 2, 2, 2)

# MindSpore
import mindspore
from mindspore import Tensor
import numpy as np

pool = mindspore.nn.MaxPool2d(kernel_size=3, stride=1)
x = Tensor(np.random.randint(0, 10, [1, 2, 4, 4]), mindspore.float32)
output = pool(x)
result = output.shape
print(result)
# (1, 2, 2, 2)

Code Example 2

Use pad mode to ensure functional consistency.

# PyTorch
import torch
import numpy as np

np_x = np.random.randint(0, 10, [1, 2, 4, 4])
x = torch.tensor(np_x, dtype=torch.float32)
max_pool = torch.nn.MaxPool2d(kernel_size=2, stride=1, padding=1, dilation=1, return_indices=False)
output = max_pool(x)
result = output.shape
print(tuple(result))
# (1, 2, 5, 5)

# MindSpore
import mindspore as ms
from mindspore import Tensor
import mindspore.nn as nn
import numpy as np

np_x = np.random.randint(0, 10, [1, 2, 4, 4])
x = Tensor(np_x, ms.float32)
max_pool = nn.MaxPool2d(kernel_size=2, stride=1, pad_mode='pad', padding=1, dilation=1, return_indices=False)
output = max_pool(x)
result = output.shape
print(result)
# (1, 2, 5, 5)