Spectral Normalization TF2

Spectral Normalization implemented as Tensorflow 2.0

TODO

• Convert simple conv2d model to DCGAN model

Run Test Code

This is currently a test code using a simple image classification model.

python main.py

Algorithm

How to use

1. Sequential API

from sn import SpectralNormalization

model = models.Sequential()
model.add(SpectralNormalization(layers.Conv2D(32, (3, 3), activation='relu')))
...

2. Functional API

from sn import SpectralNormalization

inputs = layers.Input(shape=(28,28,1))
x = SpectralNormalization(layers.Conv2D(32, (3, 3), activation='relu'))(inputs)
...

3. Custom Layer Method

from sn import SpectralNormalization

class CustomLayer(tf.keras.layers.Layer):
    def __init__(self):
        self.conv2DSN = SpectralNormalization(layers.Conv2D(32, (3, 3), activation='relu'))
        ...
    
    def call(self, inputs):
        x = self.conv2DSN(inputs)
        ...

Rerference

Spectral Normalization for Generative Adversarial Networks