MC Dropout API¶

This page documents the wrapper used to keep dropout active at inference time and summarize predictive spread.

Public objects¶

MCDropoutWrapper

Parameter and variable conventions¶

Name	Meaning
`model`	dropout-enabled base model
`n_mc`	number of stochastic forward passes
`apply_softmax`	convert logits to probabilities before aggregation

Workflow expectations¶

train the base model normally with dropout layers present
wrap it with MCDropoutWrapper(model, n_mc=..., apply_softmax=...)
call predict(...) or predict_uq(...)

Input and output shapes¶

x can have any shape accepted by the wrapped model.
predict(...) returns (mean, var) with the same trailing shape as a single forward pass.
if apply_softmax=True, the last dimension is interpreted as class probability.

`UQResult` mapping¶

predict_uq(...) populates mean, epistemic_var, and total_var. When apply_softmax=True, it additionally populates probs and probs_var.

Common preconditions and failure modes¶

the wrapped model must already contain dropout layers for MC Dropout to have any effect
n_mc should be positive and large enough to stabilize the variance estimate
apply_softmax=True should only be used when the wrapped model emits logits

Minimal example¶

mc_model = MCDropoutWrapper(model, n_mc=50, apply_softmax=False)
uq = mc_model.predict_uq(x_test)

deepuq.methods.mc_dropout ¶

Monte-Carlo dropout inference wrappers.

The wrapper in this module keeps dropout layers active at inference time, collects repeated stochastic forward passes, and summarizes their spread as a predictive uncertainty estimate.

MCDropoutWrapper ¶

Bases: Module

Wrap a dropout-enabled model to perform MC Dropout at inference.

Parameters:

Name	Type	Description	Default
`model`	`Module`	`torch.nn.Module` that already contains dropout layers.	required
`n_mc`	`int`	Number of stochastic forward passes used at prediction time.	`20`
`apply_softmax`	`bool`	If `True`, interpret model outputs as logits and return probability-space moments.	`True`

predict ¶

predict(x: Tensor)

Run stochastic dropout passes and return predictive mean/variance.

Parameters:

Name	Type	Description	Default
`x`	`Tensor`	Input batch with shape accepted by the wrapped model.	required

Returns:

Type	Description
`(mean, var):`	Tensors with the same trailing shape as a single model prediction. For classification with `apply_softmax=True`, the last dimension is class probability.

predict_uq ¶

predict_uq(x: Tensor) -> UQResult

Return predictive moments as a :class:deepuq.types.UQResult.

The wrapper reports dropout spread as epistemic_var. No explicit aleatoric component is modeled.

train ¶

train(mode: bool = True)

Mirror the wrapped model's train/eval state.

MC Dropout still forces dropout-active behavior inside predict and predict_uq by temporarily calling self.model.train(True).