比较与torch.nn.MaxPool2d的功能差异

torch.nn.MaxPool2d

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

更多内容详见torch.nn.MaxPool2d。

mindspore.nn.MaxPool2d

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

更多内容详见mindspore.nn.MaxPool2d。

使用方式

PyTorch：可以通过padding参数调整输出的shape。若输入的shape为 \( (N, C, H_{in}, W_{in}) \)，则输出的shape为 \( (N, C, H_{out}, W_{out}) \)，其中

\[ H_{out} = \left\lfloor\frac{H_{in} + 2 * \text{padding[0]} - \text{dilation[0]} \times (\text{kernel_size[0]} - 1) - 1}{\text{stride[0]}} + 1\right\rfloor \]

\[ W_{out} = \left\lfloor\frac{W_{in} + 2 * \text{padding[1]} - \text{dilation[1]} \times (\text{kernel_size[1]} - 1) - 1}{\text{stride[1]}} + 1\right\rfloor \]

MindSpore：没有padding参数，仅通过pad_mode参数控制pad模式。若输入的shape为 \( (N, C, H_{in}, W_{in}) \)，则输出的shape为 \( (N, C, H_{out}, W_{out}) \)，其中

1. pad_mode为”valid”：

   \[ H_{out} = \left\lfloor\frac{H_{in} - ({kernel\_size[0]} - 1)}{\text{stride[0]}}\right\rfloor \]
   \[ W_{out} = \left\lfloor\frac{W_{in} - ({kernel\_size[1]} - 1)}{\text{stride[1]}}\right\rfloor \]

2. pad_mode为”same”：

   \[ H_{out} = \left\lfloor\frac{H_{in}}{\text{stride[0]}}\right\rfloor \]
   \[ W_{out} = \left\lfloor\frac{W_{in}}{\text{stride[1]}}\right\rfloor \]

代码示例

import mindspore
from mindspore import Tensor
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 = 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 = 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])