Function Differences with torch.nn.MaxPool2d

torch.nn.MaxPool2d

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

For more information, see torch.nn.MaxPool2d.

mindspore.nn.MaxPool2d

class mindspore.nn.MaxPool2d(
    kernel_size=1,
    stride=1,
    pad_mode="valid",
    data_format="NCHW"
)

For more information, see mindspore.nn.MaxPool2d.

Differences

PyTorch：The output shape can be adjusted through the padding parameter. If the shape of input is $(N,C,H_{in},W_{in})$，the shape of output is $(N,C,H_{out},W_{out})$, where

$$H_{out}=\lfloor \frac{H_{in}+2\ast \text{padding[0]}-\text{dilation[0]}\times (\text{kernel\_size[0]}-1)-1}{\text{stride[0]}}+1\rfloor$$

$$W_{out}=\lfloor \frac{W_{in}+2\ast \text{padding[1]}-\text{dilation[1]}\times (\text{kernel\_size[1]}-1)-1}{\text{stride[1]}}+1\rfloor$$

MindSpore：There is no padding parameter, the pad mode is controlled by the pad_mode parameter only. If the shape of input is $(N,C,H_{in},W_{in})$，the shape of output is $(N,C,H_{out},W_{out})$, where

1. pad_mode is “valid”：

   $$H_{out}=\lceil \frac{H_{in}-(kernel\mathrm{\_}size[0]-1)}{\text{stride[0]}}\rceil$$
   $$W_{out}=\lceil \frac{W_{in}-(kernel\mathrm{\_}size[1]-1)}{\text{stride[1]}}\rceil$$

2. pad_mode is “same”：

   $$H_{out}=\lceil \frac{H_{in}}{\text{stride[0]}}\rceil$$
   $$W_{out}=\lceil \frac{W_{in}}{\text{stride[1]}}\rceil$$

Code Example

import mindspore as ms
import mindspore.nn as nn
import torch
import numpy as np

# In MindSpore, pad_mode="valid"
pool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode="valid")
input_x = ms.Tensor(np.random.randn(20, 16, 50, 32).astype(np.float32))
output = pool(input_x)
print(output.shape)
# Out：
# (20, 16, 24, 15)

# In MindSpore, pad_mode="same"
pool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode="same")
input_x = ms.Tensor(np.random.randn(20, 16, 50, 32).astype(np.float32))
output = pool(input_x)
print(output.shape)
# Out：
# (20, 16, 25, 16)


# In torch, padding=1
m = torch.nn.MaxPool2d(3, stride=2, padding=1)
input_x = torch.randn(20, 16, 50, 32)
output = m(input_x)
print(output.shape)
# Out：
# torch.Size([20, 16, 25, 16])