kgcnn.initializers package

Submodules

kgcnn.initializers.initializers module

class kgcnn.initializers.initializers.GlorotOrthogonal(gain=1.0, seed=None, scale=1.0, mode='fan_avg')

Bases: keras.src.initializers.random_initializers.OrthogonalInitializer

Combining Glorot variance and Orthogonal initializer.

Generate a weight matrix with variance according to Glorot initialization. Based on a random (semi-) orthogonal matrix neural networks are expected to learn better when features are de-correlated.

This is stated by e.g.:

• “Reducing over-fitting in deep networks by de-correlating representations” by M. Cogswell et al. (2016) https://arxiv.org/abs/1511.06068 .

• “Dropout: a simple way to prevent neural networks from over-fitting” by N. Srivastava et al. (2014) https://dl.acm.org/doi/10.5555/2627435.2670313 .

• “Exact solutions to the nonlinear dynamics of learning in deep linear neural networks” by A. M. Saxe et al. (2013) https://arxiv.org/abs/1312.6120 .

This implementation has been borrowed and slightly modified from DimeNetPP .

get_config()

Get keras config.

class kgcnn.initializers.initializers.HeOrthogonal(gain=1.0, seed=None, scale=1.0, mode='fan_in')

Bases: keras.src.initializers.random_initializers.OrthogonalInitializer

Combining He variance and Orthogonal initializer.

Generate a weight matrix with variance according to He initialization. Based on a random (semi-)orthogonal matrix neural networks are expected to learn better when features are de-correlated.

This is stated by e.g.:

• “Reducing over-fitting in deep networks by de-correlating representations” by M. Cogswell et al. (2016) https://arxiv.org/abs/1511.06068 .

• “Dropout: a simple way to prevent neural networks from over-fitting” by N. Srivastava et al. (2014)

  https://dl.acm.org/doi/10.5555/2627435.2670313 .

• “Exact solutions to the nonlinear dynamics of learning in deep linear neural networks” by A. M. Saxe et al. (2013) https://arxiv.org/abs/1312.6120 .

This implementation has been borrowed and slightly modified from GemNet .

static _standardize(kernel, shape)

Standardize kernel over fan_in dimensions.

Parameters

• kernel – Kernel variable.

• shape – Shape of the kernel.

get_config()

Get keras config.

kgcnn.initializers.initializers._compute_fans(shape)

Computes the number of input and output units for a weight shape.

Taken from original TensorFlow implementation and copied here for static reference.

Parameters

shape – Integer shape tuple or tensor shape.

Returns

A tuple of integer scalars (fan_in, fan_out).

Module contents