incerto.active.BatchBALDAcquisition#
- class incerto.active.BatchBALDAcquisition(num_samples=20)[source]#
Bases:
BaseAcquisitionApproximate BatchBALD via individual BALD scores.
Full BatchBALD (Kirsch et al., NeurIPS 2019) greedily selects batches that jointly maximise information gain by computing joint entropies. This implementation returns per-sample BALD scores as a tractable approximation; for true batch-aware selection, pair with a diversity-aware strategy (e.g.,
DiversitySampling).- Reference:
Kirsch et al., “BatchBALD: Efficient and Diverse Batch Acquisition for Deep Bayesian Active Learning” (NeurIPS 2019)
- Parameters:
num_samples (
int)
- __init__(num_samples=20)[source]#
Initialize BatchBALD acquisition.
- Parameters:
num_samples (
int) – Number of MC samples
Methods
__init__([num_samples])Initialize BatchBALD acquisition.
score(model, x, **kwargs)Compute per-sample BALD scores as a BatchBALD approximation.
- __init__(num_samples=20)[source]#
Initialize BatchBALD acquisition.
- Parameters:
num_samples (
int) – Number of MC samples
- score(model, x, **kwargs)[source]#
Compute per-sample BALD scores as a BatchBALD approximation.
Note: This is a simplified version that returns individual BALD scores. Full BatchBALD requires joint entropy computation which is computationally expensive. For batch-aware diversity, pair with a diversity-aware strategy like
DiversitySampling.