Intervention Policies¶

This module provides policies for selecting which concepts to intervene on during inference.

Summary¶

Policy Classes

`UniformPolicy`	Uniform intervention policy that assigns equal priority to all concepts.
`RandomPolicy`	Random intervention policy that generates random values for concept selection.
`UncertaintyInterventionPolicy`	Uncertainty-based intervention policy using distance from a maximum uncertainty point.

Class Documentation¶

class UniformPolicy(out_features: int)[source]¶

Bases: BaseConceptLayer

Uniform intervention policy that assigns equal priority to all concepts.

This policy returns zeros for all concepts, indicating uniform/equal uncertainty or priority across all concepts. Useful as a baseline where no concept is preferred over others.

out_features¶

Number of output features.

Type:: int

Parameters:: out_features – Number of output concept features.

Example

>>> import torch
>>> from torch_concepts.nn import UniformPolicy
>>>
>>> # Create uniform policy
>>> policy = UniformPolicy(out_features=10)
>>>
>>> # Generate random concept endogenous
>>> endogenous = torch.randn(4, 10)  # batch_size=4, n_concepts=10
>>>
>>> # Apply policy - returns zeros (uniform priority)
>>> scores = policy(endogenous)
>>> print(scores.shape)  # torch.Size([4, 10])
>>> print(torch.all(scores == 0.0))  # True
>>>
>>> # Useful for baseline comparisons
>>> # All concepts have equal intervention priority
>>> print(scores.mean())  # tensor(0.)
>>> print(scores.std())   # tensor(0.)

forward(endogenous: Tensor) → Tensor[source]¶

Generate uniform (zero) intervention scores.

Parameters:: endogenous – Input concept endogenous of shape (batch_size, n_concepts).
Returns:: Zeros tensor of same shape as input.
Return type:: torch.Tensor

training: bool¶

class RandomPolicy(out_features: int, scale: float = 1.0)[source]¶

Bases: BaseConceptLayer

Random intervention policy that generates random values for concept selection.

This policy generates random values scaled by a factor, useful for random baseline comparisons in intervention experiments.

out_features¶

Number of output features.

Type:: int

scale¶

Scaling factor for random values.

Type:: float

Parameters:

out_features – Number of output concept features.
scale – Scaling factor for random values (default: 1.0).

Example

>>> import torch
>>> from torch_concepts.nn import RandomPolicy
>>>
>>> # Create random policy
>>> policy = RandomPolicy(out_features=10, scale=2.0)
>>>
>>> # Generate random concept endogenous
>>> endogenous = torch.randn(4, 10)  # batch_size=4, n_concepts=10
>>>
>>> # Apply policy to get random intervention scores
>>> scores = policy(endogenous)
>>> print(scores.shape)  # torch.Size([4, 10])
>>> print(scores.min() >= 0.0)  # True (absolute values)
>>> print(scores.max() <= 2.0)  # True (scaled by 2.0)
>>>
>>> # Each call generates different random values
>>> scores2 = policy(endogenous)
>>> print(torch.equal(scores, scores2))  # False

forward(endogenous: Tensor) → Tensor[source]¶

Generate random intervention scores.

Parameters:: endogenous – Input concept endogenous of shape (batch_size, n_concepts).
Returns:: Random scores of same shape as input, scaled by self.scale.
Return type:: torch.Tensor

training: bool¶

class UncertaintyInterventionPolicy(out_features: int, max_uncertainty_point: float = 0.0)[source]¶

Bases: BaseConceptLayer

Uncertainty-based intervention policy using distance from a maximum uncertainty point.

This policy measures uncertainty as the distance of concept endogenous from a maximum uncertainty point. Values closer to this point are considered more uncertain, while values further from this point are considered more certain.

out_features¶

Number of output features.

Type:: int

max_uncertainty_point¶

The point where uncertainty is maximum.

Type:: float

Parameters:

out_features – Number of output concept features.
max_uncertainty_point – The value representing maximum uncertainty (default: 0.0). Values closer to this point are more uncertain, values further away are more certain.

Example

>>> import torch
>>> from torch_concepts.nn import UncertaintyInterventionPolicy
>>>
>>> # Create uncertainty policy with default max uncertainty point (0.0)
>>> policy = UncertaintyInterventionPolicy(out_features=10)
>>>
>>> # Generate concept endogenous with varying confidence
>>> endogenous = torch.tensor([
...     [3.0, -2.5, 0.1, -0.2, 4.0],  # High confidence for 1st, 2nd, 5th
...     [0.5, 0.3, -0.4, 2.0, -1.5]   # Mixed confidence
... ])
>>>
>>> # Apply policy - returns distance from max uncertainty point (certainty scores)
>>> scores = policy(endogenous)
>>> print(scores)
>>> # tensor([[3.0, 2.5, 0.1, 0.2, 4.0],
>>> #         [0.5, 0.3, 0.4, 2.0, 1.5]])
>>>
>>> # Higher scores = higher certainty = lower intervention priority
>>> # For intervention, you'd typically intervene on LOW scores
>>> print(scores[0].argmin())  # tensor(2) - most uncertain concept
>>> print(scores[0].argmax())  # tensor(4) - most certain concept
>>>
>>> # Use custom max uncertainty point (e.g., 0.5 for probabilities)
>>> policy_prob = UncertaintyInterventionPolicy(out_features=5, max_uncertainty_point=0.5)
>>> probs = torch.tensor([[0.1, 0.5, 0.9, 0.45, 0.55]])
>>> certainty = policy_prob(probs)
>>> print(certainty)
>>> # tensor([[0.4, 0.0, 0.4, 0.05, 0.05]])
>>> # Values at 0.5 are most uncertain, values at 0.1 or 0.9 are most certain

forward(endogenous: Tensor) → Tensor[source]¶

Compute certainty scores as distance from maximum uncertainty point.

Parameters:

endogenous – Input concept endogenous of shape (batch_size, n_concepts).

Returns:

Distance from max uncertainty point (certainty scores) of same shape as input.: Higher values indicate higher certainty (further from max uncertainty point). Lower values indicate higher uncertainty (closer to max uncertainty point).

Return type:

torch.Tensor

training: bool¶