/ICNR

Convolution NN resize initialization for subpixel convolutions

Primary LanguageJupyter NotebookMIT LicenseMIT

Checkerboard artifact free sub pixel convolution initialization

Tensorflow implementation of ICNR initialization used in https://arxiv.org/pdf/1707.02937.pdf

screenshot

Updated for TF v2.4.

Usage example:

Wrap up your initialization with ICRN and you are ready to go:

layers.Conv2DTranspose(..., strides=scale, kernel_initializer=ICNR(GlorotUniform(), scale))