Function Differences with torch.nn.MaxPool3D

View Source On Gitee

torch.nn.MaxPool3D

torch.nn.MaxPool3D(
    kernel_size=1,
    stride=None,
    padding=0,
    dilation=1,
    return_indices=False,
    ceil_mode=False
)

For more information, see torch.nn.MaxPool3D.

mindspore.ops.MaxPool3D

class mindspore.ops.MaxPool3D(
    kernel_size=1,
    strides=1,
    pad_mode='valid',
    pad_list=0,
    ceil_mode=None,
    data_format='NCDHW'
)(input)

For more information, see mindspore.ops.MaxPool3D.

Differences

PyTorch: Supports both 5-dimensional (N, C, Din, Hin, Win) input data and 4-dimensional (C, Din, Hin, Win) input data.

MindSpore: Supports only 5-dimensional (N, C, Din, Hin, Win) input data.

Migration advice: If you need MindSpore MaxPool3D to calculate on 4-dimensional input data, data can be expanded to 5-dimensional using ExpandDims operator and passed into MaxPool3D. You can then use the Squeeze operator to convert the dimension from (1, C, Dout, Hout, Wout) to (C, Dout, Hout, Wout).

Code Example

import mindspore
from mindspore import Tensor
import mindspore.ops as ops
import torch
import numpy as np

# In MindSpore
net = ops.MaxPool3D((3, 2, 2), strides=2)
x = Tensor(np.ones([20, 16, 50, 44, 31]), mindspore.float32)
output = net(x).shape
print(output)
# Out:
# (20, 16, 24, 22, 15)

# In PyTorch
m = torch.nn.MaxPool3d((3, 2, 2), stride=2)
input = torch.rand(20, 16, 50, 44, 31)
output = m(input).shape
print(output)
# Out:
# torch.Size([20, 16, 24, 22, 15])