tfwn

Weight Normalization layer wrapper for TensorFlow-Keras API.

Inspired by Sean Morgan implementation, but:

• No data initialization (only eager mode was implemented in original pull request).
• Code refactoring
• More tests
• CIFAR10 example from original paper reimplemented

Examples

Unfortunately I couldn't reproduce parer results on CIFAR10 with batch size 100. As you can see there is no much difference in accuracy.

But with much smaller batch size model with weight normalization is much better then regular one.

How to use

import tensorflow as tf
from tfwn import WeightNorm


dense_wn = WeightNorm(tf.keras.layers.Dense(3))
out = dense_wn(input)

References

Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks

Tim Salimans, and Diederik P. Kingma.

@inproceedings{Salimans2016WeightNorm,
  title={Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks},
  author={Tim Salimans and Diederik P. Kingma},
  booktitle={Neural Information Processing Systems 2016},
  year={2016}
}