Instance Normalization

This repo implements the Instance Normalization with different programming languages or libraries: Tensorflow (Keras), Numpy, C++, Eigen, CUDA. Each sample includes both the forward and backward passes and results are checked compared with Keras's InstanceNormalization Instance.

Instance Normalization Forward Pass

Suppose $N, C, D$ are batch, channel and feature dimensions respectively.

$x: (N, C, D)$

$\mu, \sigma ^2: (N, C, 1)$

$\gamma, \beta: (1, C, 1)$

$\mu = \frac{1}{D}\sum_{i=1}^{D}x_i$

$\sigma^2 = \frac{1}{D}\sum_{i=1}^{D}(x_i - \mu)^2$

$\hat{x} = \frac{x - \mu}{\sqrt{\sigma^2 + \epsilon}}$

$y=\gamma\hat{x} + \beta$

Instance Normalization Backward Pass

$dy: (N, C, D)$

$d\mu, d\sigma ^2: (N, C, 1)$

$d\gamma, d\beta: (1, C, 1)$

$d\gamma=dy\cdot \frac{\partial y}{\partial \gamma}=\sum_{i=1}^{ND}dy\cdot\frac{x-\mu}{\sqrt{\sigma^2 + \epsilon}}$

$d\beta=dy\cdot \frac{\partial y}{\partial \beta}=\sum_{i=1}^{ND}dy$

Here we can view $y=\gamma\hat{x} + \beta$ as a composition function $y(i(x), \sigma^2(x), \mu(x))$ , where $i(x)=x$ and $\sigma^2(x)$ , $\mu(x)$ are defined as above. Therefore, we get:

$dx = \frac{\partial L}{\partial i}\cdot\frac{\partial i}{\partial x} + \frac{\partial L}{\partial \sigma^2}\cdot\frac{\partial \sigma^2}{\partial x} + \frac{\partial L}{\partial \mu}\cdot\frac{\partial \mu}{\partial x}$

Compute $\frac{\partial L}{\partial i}\cdot\frac{\partial i}{\partial x}$

$\frac{\partial L}{\partial i} = \frac{\partial L}{\partial y}\cdot\frac{\partial y}{\partial i} = dy \cdot \gamma \cdot \frac{1}{\sqrt{\sigma^2 + \epsilon}}$

$\frac{\partial i}{\partial x}=1$

Compute $\frac{\partial L}{\partial \sigma^2}\cdot\frac{\partial \sigma^2}{\partial x}$

$\frac{\partial L}{\partial \sigma^2} = \frac{\partial L}{\partial y}\cdot\frac{\partial y}{\partial \sigma^2} = \sum_{i=1}^{D}dy \cdot \gamma (x-\mu)(-\frac{1}{2})(\sigma^2 + \epsilon)^{-\frac{3}{2}}$

$\frac{\partial \sigma^2}{\partial x}=\frac{2}{D}(x-\mu)$

Compute $\frac{\partial L}{\partial \mu}\cdot\frac{\partial \mu}{\partial x}$

$\frac{\partial L}{\partial \mu}=\frac{\partial L}{\partial \mu}\cdot \frac{\partial \mu}{\partial \mu} + \frac{\partial L}{\partial \sigma^2}\cdot \frac{\partial \sigma^2}{\partial \mu}\\=\frac{\partial L}{\partial y}\cdot \frac{\partial y}{\partial \mu}\cdot 1 + \frac{\partial L}{\partial \sigma^2}\cdot \frac{\partial \sigma^2}{\partial \mu}\\=\sum_{i=1}^{D}dy\cdot \gamma \cdot\frac{-1}{\sqrt{\sigma^2+\epsilon}}+\sum_{i=1}^{D}\frac{\partial L}{\partial \sigma^2}\cdot \frac{1}{D}(-2)(x-\mu)$ , where the last term is 0

$\frac{\partial \mu}{\partial x}=\frac{1}{D}$

Name		Name	Last commit message	Last commit date
Latest commit History 13 Commits
cub		cub
Makefile		Makefile
README.md		README.md
instance_norm.cpp		instance_norm.cpp
instance_norm.cu		instance_norm.cu
instance_norm_cpp.py		instance_norm_cpp.py
instance_norm_eigen.cpp		instance_norm_eigen.cpp
instance_norm_np.py		instance_norm_np.py
instance_norm_tf.py		instance_norm_tf.py
runme.sh		runme.sh

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

Instance Normalization

Instance Normalization Forward Pass

Instance Normalization Backward Pass

About

Uh oh!

Releases

Packages

Languages

crapromer/instance_norm

Folders and files

Latest commit

History

Repository files navigation

Instance Normalization

Instance Normalization Forward Pass

Instance Normalization Backward Pass

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages