src.model.deeplearn.layer.dense_with_ground_row

Classes

DenseWithGroundRow(*args, **kwargs)

class src.model.deeplearn.layer.dense_with_ground_row.DenseWithGroundRow(*args, **kwargs)
Author:

Alberto M. Esmoris Pena

Fused “strip-ground-row → Dense → prepend-ground-row” block for the SpConv architecture.

The SpConv architecture stores per-depth feature tensors in padded form (an extra ground row at index 0 holds zeros). A Dense projection should only be applied to the active-cell view; the ground row must remain all zeros so the next tf.gather(F, S[1:]) step still fetches zero features for “missing neighbor” entries (S[v p] == 0).

The layer owns the Dense kernel + bias directly so a single graph node implements the slice + dense + pad sequence:

  • one node per call site (instead of three Lambda + Dense layers wired functionally);

  • model.summary() is readable (one entry per dense site);

  • the optional activation is applied inline without a separate Activation layer wrapping;

  • get_config / from_config round-trip is straightforward (no Lambda lambdas to serialize).

The math is: matmul on the active-cell view, optional bias, optional activation, then re-prepend the ground row at index 0.

Variables:

  • units (int) – Number of output channels (same as Keras keras.layers.Dense).

  • activation – Optional activation applied before prepending the ground row (so the ground row stays at zero regardless of bias / activation choice).

__init__(units, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs)

Initialize the member attributes of the layer and the internal weights that do not depend on the input dimensionality.

Parameters:

kwargs – The key-word specification to parametrize the layer.

build(input_shape)

Logic to build the layer before the first call is executed.

This method can be overloaded by any derived class to either change or extend the logic.

Parameters:

dim_in – The dimensionality of the input tensor.

call(inputs, training=False, mask=False)
Parameters:

inputs – Padded sparse-conv tensor with shape (1 + R, n_in). Row 0 is the ground row.

Returns:

Padded tensor with shape (1 + R, units). Row 0 is the new ground row (all zeros); rows 1..R are activation(W · inputs[1:] + b).

compute_output_shape(input_shape)
get_config()

Obtain the dictionary specifying how to serialize the layer.

Returns:

The dictionary with the necessary data to serialize the layer.

Return type:

dict