Constraints¶

Setting Defaults¶

Since constraints are often the same for all layers in an SPN, libspn-keras provides the following functions to get and set default constraints. These can still be overridden by providing the initializers explicitly at initialization of a layer.

Linear accumulator constraints¶

By default, the logspace accumulator constraint is set to GreaterEqualEpsilonNormalized.

libspn_keras.set_default_linear_accumulators_constraint(op)¶

Set default sum op to conveniently use it throughout an SPN architecture.

Parameters: op (Constraint) – A constraint applied to linear accumulators after updates
Return type: None

libspn_keras.get_default_linear_accumulators_constraint()¶

Get default linear accumulator constraint.

Return type: Optional[Constraint]
Returns: A Constraint instance that was set with set_default_linear_accumulators_constraint

Logspace accumulator constraints¶

By default, the logspace accumulator constraint is set to LogNormalized.

libspn_keras.set_default_logspace_accumulators_constraint(op)¶

Set default sum op to conveniently use it throughout an SPN architecture.

Parameters: op (Constraint) – A constraint applied to logspace accumulators after updates
Return type: None

libspn_keras.get_default_logspace_accumulators_constraint()¶

Get default logspace accumulator constraint.

Return type: Optional[Constraint]
Returns: A Constraint instance that was set with set_default_logspace_accumulators_constraint

Linear Weight Constraints¶

These should be used for linear accumulators.

class libspn_keras.constraints.GreaterEqualEpsilonNormalized(epsilon=1e-10, axis=- 2)¶

Constraints the weight to be greater than or equal to epsilon and then normalizes.

Parameters: epsilon (float) – Constant, usually small non-zero

class libspn_keras.constraints.GreaterEqualEpsilon(epsilon=1e-10)¶

Constraints the weight to be greater than or equal to epsilon.

Parameters: epsilon (float) – Constant, usually small non-zero

Log Weight Constraints¶

These should be used for log accumulators.

class libspn_keras.constraints.LogNormalized(axis=- 2)¶

Normalizes log-space weights.

Parameters: axis (int) – Axis along whichto normalize

Scale Constraints¶

The following constraint is useful for ensuring stable scale parameters in location-scale leaf layers.

class libspn_keras.constraints.Clip(min, max=None)¶

Constraints the weights to be between min and max.

Parameters

min (float) – Minimum clip value
max (Optional[float]) – Maximum clip value