Skip to content




Parent class: BatchNormND

Derived classes: -

This module implements the batch normalization operation for five-dimensional tensors of shape (batchsize, maps, d, h, w), for example, after 3-dimensional convolutional layers. For more detailed theoretical information see . BatchNormND.


def __init__(self, maps, epsilon=1e-5, initFactor=1.0, minFactor=0.1, sscale=0.01, affine=True, name=None, empty=False, inplace=False):


Parameter Allowed types Description Default
size int Number of input features -
epsilon float Stabilizing constant 1e-5
initFactor float Initial factor value in the moving average 1.0
minFactor float Minimal factor value in the moving average 0.1
sscale float Dispersion of the Gaussian distribution for the scale parameter of batch normalization 0.01
affine bool If True, the layer will have trainable affine parameters scale and bias True
name str Layer name None
empty bool If True, the tensors of the parameters of the module will not be initialized False
inplace bool If True, the output tensor will be written in memory in the place of the input tensor False



In order to understand how initFactor and minFactor are used, it should be mentioned that the principle of the moving average is used in the calculation of statistical parameters in the module, i.e.: \begin{equation} \hat{\mu} = \alpha\hat{\mu} + (1 - \alpha)\mu \end{equation} \begin{equation} \hat{\sigma}^2 = \alpha\hat{\sigma}^2 + (1 - \alpha)\sigma^2 \end{equation}


\hat{\mu}, \mu - moving average and average on the batch, respectively;
\hat{\sigma^2}, \sigma^2 - moving variance and variance on the batch, respectively; \alpha - conservation factor.

\alpha in the module is calculated as follows: $$ \alpha = max(\frac{IF}{n}, MF) $$


IF - initFactor, MF - minFactor, n - batch number.

size - number of input features, i.e. the maps axis of the tensor of shape (batchsize, maps, d, h, w);

epsilon - small number added to prevent division by zero during the normalization of features (see the theory in BatchNormND;

affine - flag that is responsible for whether the scale and bias parameters of the batch normalization layer will be trained, or they will be fixed (1 and 0, respectively), so that the layer will only perform normalization by the average and the variance;

inplace - flag showing whether additional memory resources should be allocated for the result. If True, then the output tensor will be written to the place of the input one in memory, which can negatively affect the network if the input tensor should take part in calculations on other branches of the graph.


Necessary imports.

import numpy as np
from PuzzleLib.Backend import gpuarray
from PuzzleLib.Modules import BatchNorm3D


gpuarray is required to properly place the tensor in the GPU.

For this module, showing tensors on the screen will not bring clarity, therefore, only the code will be given in this example. For visualized examples, please see BatchNorm or BatchNorm1D

Let us create a synthetic data tensor:

batchsize, maps, d, h, w = 5, 3, 10, 10, 10
data = gpuarray.to_gpu(np.arange(batchsize * maps * d * h * w).reshape(batchsize, maps, d, h, w).astype(np.float32))

We initialize the class object with default parameters and apply it to the data:

bn = BatchNorm3D(maps)

We can see the calculated mean and variance:


As well as the initialized scales and default shifts (the scales are randomly normally distributed, the shifts are zero):