Differences with torch.nn.InstanceNorm2d

torch.nn.InstanceNorm2d

class torch.nn.InstanceNorm2d(
    num_features,
    eps=1e-05,
    momentum=0.1,
    affine=False,
    track_running_stats=False
)(input) -> Tensor

For more information, see torch.nn.InstanceNorm2d.

mindspore.nn.InstanceNorm2d

class mindspore.nn.InstanceNorm2d(
    num_features,
    eps=1e-5,
    momentum=0.1,
    affine=True,
    gamma_init='ones',
    beta_init='zeros'
)(x) -> Tensor

For more information, see mindspore.nn.InstanceNorm2d.

Differences

PyTorch: Apply normalization within each channel of the four-dimension input (2D with additional mini-batch and channel channels).

MindSpore: MindSpore API implements the same function as PyTorch, with two typical differences. The default value of the affine parameter in MindSpore is True, which learns the internal parameters γ and β, and the default value of PyTorch is False, which does not perform parameter learning. PyTorch supports the track_running_stats parameter. If set to True, it will use the mean and variance obtained from training in inference, and the default value is False. MindSpore does not have this parameter, and will use the computed mean and variance of the input data in both training and inference, with the same behavior as PyTorch default value.

Categories	Subcategories	PyTorch	MindSpore	Difference
Input	Single input	input	x	Interface input, same function, only different parameter names
Parameters	Parameter 1	num_features	num_features	-
	Parameter 2	eps	eps	-
	Parameter 3	momentum	momentum	-
	Parameter 4	affine	affine	The default values are different: MindSpore defaults to True, which learns the internal parameters γ and β, and PyTorch defaults to False, which does not learn the parameters
	Parameter 5	track_running_stats	-	If set to True, PyTorch will use the mean and variance obtained from training in inference, and the default value is False. MindSpore does not have this parameter, and will use the computed mean and variance of the input data in both training and inference, with the same behavior as PyTorch default value.
	Parameter 6	-	gamma_init	Initialize transform parameter γ for learning, default is ‘ones’, while PyTorch can’t set additionally, only ‘ones’
	Parameter 7	-	beta_init	Initialize transform parameter γ for learning, default is ‘zeros’, while PyTorch can’t set additionally, only ‘zeros’

Code Example

MindSpore affine, when set to False, has the same functions as PyTorch default behavior.

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

m = nn.InstanceNorm2d(num_features=3)
input_x = tensor(np.array([[[[0.1, 0.2], [0.3, 0.4]],
                          [[0.5, 0.6], [0.7, 0.8]],
                          [[0.9, 1], [1.1, 1.2]]]]).astype(np.float32))
output = m(input_x)
print(output.detach().numpy())
# [[[[-1.3411044  -0.44703478]
#    [ 0.4470349   1.3411044 ]]
#
#   [[-1.3411045  -0.44703463]
#    [ 0.44703472  1.3411046 ]]
#
#   [[-1.3411034  -0.44703388]
#    [ 0.44703573  1.3411053 ]]]]

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

m = nn.InstanceNorm2d(num_features=3, affine=False)
m.set_train()
input_x = Tensor(np.array([[[[0.1, 0.2], [0.3, 0.4]],
                          [[0.5, 0.6], [0.7, 0.8]],
                          [[0.9, 1], [1.1, 1.2]]]]).astype(np.float32))
output = m(input_x)
print(output)
# [[[[-1.3411045  -0.4470348 ]
#    [ 0.44703496  1.3411045 ]]
#
#   [[-1.341105   -0.4470351 ]
#    [ 0.44703424  1.3411041 ]]
#
#   [[-1.3411034  -0.44703388]
#    [ 0.44703573  1.3411053 ]]]]