Does TensorFlow provide a half normal initialiser?

[nbro]

System information

• TensorFlow version (you are using): 2.
• Are you willing to contribute it (Yes/No): maybe

Describe the feature and the current behavior/state.

I would like to have access to more initializers for variables (not just the weights of layers), but also variables e.g. created with add_weight. Specifically, I would like to have an initializer for half normals, i.e. normals that produce only positive numbers (it's like applying the absolute value to a normal).

Will this change the current api? How?

No.

Who will benefit with this feature?

Everyone.

Any Other info.

I've asked a question on Stack Overflow regarding my specific problem that TF currently doesn't seem to address.

[nbro]

Here's a custom initialiser that does what I want.

import tensorflow as tf


def random_half_normal(shape, **kwargs):
    return tf.abs(tf.keras.backend.random_normal(shape, **kwargs))


class MyLayer(tf.keras.layers.Layer):
    def build(self, input_shape):
        self.my_var = self.add_weight(initializer=random_half_normal, 
                                      trainable=False)

    def call(self, inputs):
        tf.print("\nself.my_var =", self.my_var)
        return inputs


def get_model():
    inp = tf.keras.layers.Input(shape=(1,))
    out = MyLayer(8)(inp)
    model = tf.keras.Model(inputs=inp, outputs=out)
    model.summary()
    return model


def train():
    model = get_model()
    model.compile(optimizer="adam", loss="mae")
    x_train = [2, 3, 4, 1, 2, 6]
    y_train = [1, 0, 1, 0, 1, 1]
    model.fit(x_train, y_train)


if __name__ == '__main__':
    train()

See also my answer on Stack Overflow.