Contains all the work done on the Instance Norm layer.
-
Comparing batchnorm and instance norm for a single image
Batch Norm (white) is almost perfectly eclipsed by Instance Norm (magenta). -
Deriving the equations for the BatchNorm/Instance Norm function.
Thank you CS231n.
Visualizing the calculations of mean and variance for Instance Norm.
Normalisation_Visualisations.ipynb- to recreate the visuals in blog post 1.Instance_Norm_as_BatchNorm_wrapper.ipynb- comparison between PyTorch/numpy and mlpack/armadillo implementations of the layer.test.cpp- An initial approach I tried for writing the layer as an independent unit and not as a wrapper of BatchNorm.bn.cpp- The final implementation chosen where the layer is created as a wrapper of the BatchNorm class.
To run the .cpp files, you need to have armadillo installed in your system.
g++ test.cpp -o test -larmadillo && ./test
g++ bn.cpp -o bn -larmadillo && ./testTo run Normalisation_Visualisations.ipynb, you would need pytorch,
torchvision, matplotlib, seaborn, numpy, PIL python libraries.
To run Instance_Norm_as_BatchNorm_wrapper.ipynb, you would need pytorch,
numpy and armadillo. To avoid issues with linking the libraries correctly, it is
better to run the notebook in Google Colab.