Differences with torch.nn.BatchNorm3d

torch.nn.BatchNorm3d

class torch.nn.BatchNorm3d(
    num_features,
    eps=1e-05,
    momentum=0.1,
    affine=True,
    track_running_stats=True
)(input) -> Tensor

For more information, see torch.nn.BatchNorm3d.

mindspore.nn.BatchNorm3d

class mindspore.nn.BatchNorm3d(
    num_features,
    eps=1e-5,
    momentum=0.9,
    affine=True,
    gamma_init='ones',
    beta_init='zeros',
    moving_mean_init='zeros',
    moving_var_init='ones',
    use_batch_statistics=None
)(x) -> Tensor

For more information, see mindspore.nn.BatchNorm3d.

Differences

PyTorch: Apply batch normalization on five-dimensional inputs (three-dimensional input with additional mini-batch and channel channels) to avoid internal covariate bias.

MindSpore：The function of this API is basically the same as that of PyTorch, with two typical differences. The default value of the momentum parameter in MindSpore is 0.9, and the momentum conversion relationship with PyTorch is 1-momentum. The behavior of the default value is the same as that of PyTorch. The parameter update strategy during training and inference is different from that of PyTorch.

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	num_features	num_features	-
	Parameter 2	eps	eps	-
	Parameter 3	momentum	momentum	The function is the same, but the default value in PyTorch is 0.1, and in MindSpore is 0.9, the conversion relationship with PyTorch’s momentum is 1-momentum, and the default value behavior is the same as PyTorch
	Parameter 4	affine	affine	-
	Parameter 5	track_running_stats	use_batch_statistics	The function is the same, and different values correspond to different default methods.
	Parameter 6	-	gamma_init	The initialization method of the γ parameter, default value: “ones”.
	Parameter 7	-	beta_init	The initialization method of the β parameter, default value: “zeros”.
	Parameter 8	-	moving_mean_init	Initialization method of dynamic average, default value: “zeros”.
	Parameter 9	-	moving_var_init	Initialization method of dynamic variance, default value: “ones”.
Input	Single input	input	x	Interface input, same function, only different parameter names

The detailed differences are as follows: BatchNorm is a special regularization method in the CV field. It has different computation processes during training and inference and is usually controlled by operator attributes. BatchNorm of MindSpore and PyTorch uses two different parameter groups at this point.

Difference 1

torch.nn.BatchNorm3d status under different parameters

training	track_running_stats	Status
True	True	Expected training status. `running_mean` and `running_var` trace the statistical features of the batch in the entire training process. Each group of input data is normalized based on the mean and var statistical features of the current batch, and then `running_mean` and `running_var` are updated.
True	False	Each group of input data is normalized based on the statistics feature of the current batch, but the `running_mean` and `running_var` parameters do not exist.
False	True	Expected inference status. The BN uses `running_mean` and `running_var` for normalization and does not update them.
False	False	The effect is the same as that of the second status. The only difference is that this is the inference status and does not learn the weight and bias parameters. Generally, this status is not used.

mindspore.nn.BatchNorm3d status under different parameters

use_batch_statistics	Status
True	Expected training status. `moving_mean` and `moving_var` trace the statistical features of the batch in the entire training process. Each group of input data is normalized based on the mean and var statistical features of the current batch, and then `moving_mean` and `moving_var` are updated.
Fasle	Expected inference status. The BN uses `moving_mean` and `moving_var` for normalization and does not update them.
None	`use_batch_statistics` is automatically set. For training, set `use_batch_statistics` to `True`. For inference, set `use_batch_statistics` to `False`.

Compared with torch.nn.BatchNorm3d, mindspore.nn.BatchNorm3d does not have two redundant states and retains only the most commonly used training and inference states.

Difference 2

In PyTorch, the network is in training mode by default, while in MindSpore, it is in inference mode by default (is_training is False). You need to use the net.set_train() method in MindSpore to switch the network to training mode. In this case, the parameters mean and variance are calculated during the training. Otherwise, in inference mode, the parameters are loaded from the checkpoint.
Difference 3

The meaning of the momentum parameter of the BatchNorm series operators in MindSpore is opposite to that in PyTorch. The relationship is as follows:

\[momentum_{pytorch} = 1 - momentum_{mindspore}\]

Code Example

In PyTorch, the value after 1-momentum is equal to the momentum of MindSpore, both trained by using mini-batch data and learning parameters.

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

m = nn.BatchNorm3d(num_features=2, momentum=0.1)
input_x = tensor(np.array([[[[[0.1, 0.2], [0.3, 0.4]]],
                             [[[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 ]]]
#
#
#   [[[-1.3411034  -0.44703388]
#     [ 0.44703573  1.3411053 ]]]]]

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

m = nn.BatchNorm3d(num_features=2, momentum=0.9)
m.set_train()
# BatchNorm3d<
#      (bn2d): BatchNorm2d<num_features=2, eps=1e-05, momentum=0.9, gamma=Parameter (name=bn2d.gamma, shape=(2,), dtype=Float32, requires_grad=True), beta=Parameter (name=bn2d.beta, shape=(2,), dtype=Float32, requires_grad=True), moving_mean=Parameter (name=bn2d.moving_mean, shape=(2,), dtype=Float32, requires_grad=False), moving_variance=Parameter (name=bn2d.moving_variance, shape=(2,), dtype=Float32, requires_grad=False)>
#      >
input_x = Tensor(np.array([[[[[0.1, 0.2], [0.3, 0.4]]],
                             [[[0.9, 1], [1.1, 1.2]]]]]).astype(np.float32))
output = m(input_x)
print(output)
# [[[[[-1.3411044  -0.44703478]
#     [ 0.4470349   1.3411044 ]]]
#
#
#   [[[-1.3411039  -0.44703427]
#     [ 0.44703534  1.341105  ]]]]]