box_embeddings.modules.intersection.tf_gumbel_intersection
¶
Module Contents¶
- tf_gumbel_intersection(left: box_embeddings.parameterizations.TFTBoxTensor, right: box_embeddings.parameterizations.TFTBoxTensor, gumbel_beta: float = 1.0, approximation_mode: Optional[str] = None) box_embeddings.parameterizations.TFTBoxTensor ¶
Hard Intersection operation as a function.
Note
This function can give fipped boxes, i.e. where z[i] > Z[i]
- Parameters
left – BoxTensor which is the left operand
right – BoxTensor which is the right operand
gumbel_beta – Beta parameter
approximation_mode – Use hard clipping (‘clipping’) or hard clipping with separeate value for forward and backward passes (‘clipping_forward’) to satisfy the required inequalities. Set None to not use any approximation. (default: None)
- Returns
The resulting BoxTensor obtained by interection. It has the same concrete type as the self (left operand).
- Raises
ValueError – When approximation_mode is not in [None, ‘clipping’, ‘clipping_forward’]
- class TFGumbelIntersection(beta: float = 1.0, approximation_mode: Optional[str] = None)¶
Bases:
box_embeddings.modules.intersection._tf_intersection._TFIntersection
Gumbel intersection operation as a Layer/Module.
Performs the intersection operation as described in Improving Local Identifiability in Probabilistic Box Embeddings .