MC-dropout predictive draws from a fine-tuned Pillar 4 fit

Runs n_draws forward passes through a fitted edaphos_foundation_classifier or edaphos_foundation_regressor with the MLP head's dropout kept in train mode, producing a Monte-Carlo sample of the predictive posterior (Gal & Ghahramani 2016). The fit must have been trained with an MLP head and a non-zero dropout for the draws to be non-degenerate.

Usage

foundation_mcdropout_predict(object, x, n_draws = 50L, seed = NULL)

Arguments

object

An edaphos_foundation_classifier or edaphos_foundation_regressor fit.

x

New patches (N, C, H, W).

n_draws

Integer; number of MC forward passes.

seed

Optional integer seed.

Value

For regression, a (n_draws, N) numeric matrix. For classification, a (n_draws, N, n_classes) array of softmax probabilities.

References

Gal, Y. and Ghahramani, Z. (2016). Dropout as a Bayesian approximation: representing model uncertainty in deep learning. ICML 33, 1050-1059.