Implementation of (https://arxiv.org/pdf/1511.06856v2.pdf) in tensorflow.
Helps initialize weights for big nets so they can train without things like batch normalization.
Not very well formatted but I thought I'd share it. Included is some sample code on how to use it.
Requires tensorflow, sklearn, re, and numpy
run python magic_init.py to test it out. Look at the code under the "if name == 'main'" block to see how to use it