Differences with torch.nn.functional.interpolate

The following mapping relationships can be found in this file.

PyTorch APIs	MindSpore APIs
torch.nn.functional.interpolate	mindspore.ops.interpolate
torch.nn.functional.upsample	mindspore.ops.upsample
torch.nn.Upsample	mindspore.nn.Upsample

torch.nn.functional.interpolate

torch.nn.functional.interpolate(
    input,
    size=None,
    scale_factor=None,
    mode='nearest',
    align_corners=None,
    recompute_scale_factor=None) -> Tensor

For more information, see torch.nn.functional.interpolate.

mindspore.ops.interpolate

mindspore.ops.interpolate(
    x,
    size=None,
    scale_factor=None,
    mode='nearest',
    align_corners=None,
    recompute_scale_factor=None) -> Tensor

For more information, see mindspore.ops.interpolate.

Usage

PyTorch: Data is upsampled or downsampled based on size or scale_factor. The recompute_scale_factor controls whether the scale_factor used for interpolation calculation is re-calculated. If recompute_scale_factor is True, scale_factor must be passed in and the output size is calculated using scale_factor. The calculated output size will be used to infer the new ratio for interpolation. When scale_factor is a floating point number, it may be different from the re-calculated ratio due to rounding and precision issues. If recompute_scale_factor is False, interpolation is performed directly using size or scale_factor. Interpolation can be performed using one of six modes: ‘nearest’ | ‘linear’ | ‘bilinear’ | ‘bicubic’ | ‘trilinear’ | ‘area’. The align_corners controls the alignment of the coordinates and is effective for ‘linear’ | ‘bilinear’ | ‘bicubic’ | ‘trilinear’ modes, with a default value of False.

MindSpore: The functionality is basically the same as PyTorch, but support for some parameters is not complete, such as some modes cannot directly pass in scale_factor, but can be circumvented by setting the recompute_scale_factor parameter to True (when scale_factor is a floating point number, accuracy errors may occur), and the specific differences are as follows.

Categories	Subcategories	PyTorch	MindSpore	Difference
Parameter	Parameter 1	size	size	-
	Parameter 2	scale_factor	scale_factor	Function is consistent. Currently only supports ‘area’ mode directly pass in `scale_factor`. For unsupported modes, you can bypass by setting `recompute_scale_factor` parameter to True (when `scale_factor` is a floating-point number, there may be precision errors)
	Parameter 3	mode	mode	Function is consistent
	Parameter 4	align_corners	align_corners	Function is consistent, but in ‘bicubic’ mode `align_corners=False`, the calculation method is the same as TensorFlow, and the results are different from PyTorch
	Parameter 5	recompute_scale_factor	recompute_scale_factor	-
Input	Single input	input	x	Same function, different parameter names

Difference Analysis and Examples

Code Example 1

Using the default ‘nearest’ mode interpolation, pass size in and the two APIs achieve the same function.

# Pytorch
import torch
import numpy as np

x = torch.tensor(np.array([[[1, 2, 3], [4, 5, 6]]]).astype(np.float32))
output = torch.nn.functional.interpolate(input=x, size=6)
print(output.numpy())
# [[[1. 1. 2. 2. 3. 3.]
#   [4. 4. 5. 5. 6. 6.]]]

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

x = Tensor(np.array([[[1, 2, 3], [4, 5, 6]]]).astype(np.float32))
output = ops.interpolate(x, size=6, mode="nearest")
print(output)
# [[[1. 1. 2. 2. 3. 3.]
#   [4. 4. 5. 5. 6. 6.]]]

Code Example 2

Using the ‘bilinear’ mode, scale with scale_factor. This mode is not directly supported by MindSpore, but the error can be avoided by setting the recompute_scale_factor parameter to True (when scale_factor is a floating point, there may be some inaccuracies).

# Pytorch
import torch
import numpy as np

x = torch.tensor(np.array([[[[1, 2, 3], [4, 5, 6]]]]).astype(np.float32))
output = torch.nn.functional.interpolate(input=x, scale_factor=2.0, mode="bilinear", align_corners=True)
print(output.numpy())
# [[[[1.        1.4000001 1.8       2.2       2.6       3.       ]
#    [2.        2.4       2.8       3.1999998 3.6000001 4.       ]
#    [3.        3.4000003 3.8       4.2000003 4.6       5.       ]
#    [4.        4.4       4.8       5.2       5.6       6.       ]]]]

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

x = Tensor(np.array([[[[1, 2, 3], [4, 5, 6]]]]).astype(np.float32))
output = ops.interpolate(x, scale_factor=2.0, recompute_scale_factor=True, mode="bilinear", align_corners=True)
print(output)
# [[[[1.        1.4       1.8       2.2       2.6       3.       ]
#    [2.        2.4       2.8000002 3.2       3.6       4.       ]
#    [3.        3.4       3.8000002 4.2       4.6       5.       ]
#    [4.        4.4       4.8       5.2       5.6       6.       ]]]]