MC-dropout predictive draws from a ConvLSTM fit

Runs n_draws forward passes through a fitted edaphos_temporal_convlstm with dropout active in train mode, producing a Monte-Carlo sample of the predictive posterior (Gal and Ghahramani 2016). The fit must have been trained with dropout_p > 0 for this to give a non-degenerate sample; when the fit has no dropout layers the draws collapse to the deterministic forward pass.

Usage

temporal_convlstm_mcdropout_predict(
  object,
  sequence,
  n_draws = 50L,
  return_sequence = NULL,
  seed = NULL
)

Arguments

object

An edaphos_temporal_convlstm whose underlying nn_module contains nn_dropout2d layers. Fits produced by temporal_convlstm_fit() with the default dropout_p = 0 work but produce a degenerate constant draw.

sequence

5-D input tensor (batch, T, C, H, W) or equivalent R array.

n_draws

Integer; number of MC forward passes.

return_sequence

Logical override; defaults to the value used at training time.

seed

Optional integer seed.

Value

A numeric array; first axis is the draw axis. Shape: (n_draws, batch, T, H, W) when return_sequence = TRUE, else (n_draws, batch, H, W).

References

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