SpectralNormalization

Implementation of Spectral Normalization for TensorFlow keras API. This layers are available on Distribute Strategy (ex. TPU).

This is a modification of TensorFlow keras layers and code is derived from TensorFlow ver1.14 (under Apache License 2.0).

Features

TensorFlow keras layers class with spectral normalization:

• SNDense
• SNConv1D
• SNConv2D
• SNConv3D

Requirement

• tensorflow >= 1.14
• numpy (only used in test code)

Usage

Replace keras layers (Conv2D etc) with spectral normalization layers as below:

from SpectralNormalization.layers import SNConv2D

# 2D convolution with spectral normalization
outputs = SNConv2D(64, (3, 3), padding='same')(inputs)

You can set singular_vector_initializer and power_iter arguments, which affect singular value estimation in spectral normalization. And also, you can use the same arguments as original keras layers.

Tests

You can test singular vector estimation, gradient calculation and singular vector assignment:

$cd layers
$python test.py

Also you can test them on colab TPU:

$cd layers
$python test.py --use_tpu

License

Apache License 2.0