dlsia.core.networks package

Submodules

dlsia.core.conformalize.conformalize_segmentation module

dlsia.core.conformalize.conformalize_segmentation.build_conformalizer_classify(model, testloader, alpha=0.1, missing_label=-1, device='cuda:0', norma=True)

Given a model, some test data, a threshold alpha, build me a conformalizer object

Parameters:

  • model (a neural network) – the input neural network

  • testloader (pytorch data loader) – pytorch data loader

  • alpha (float) – the level alpha

  • missing_label (int, typically -1) – missing label - data can be annotated sparsely

  • device ('cpu' / 'cuda:0') – where do we calculate things

  • norma (bool) – does the network return a normalized score? If False, run an additonal softmax

Returns:

a conformalize_classification object

Return type:

conformalize_classification

class dlsia.core.conformalize.conformalize_segmentation.conformalize_classification(alpha, estimated_label_probabilities, true_labels)

Bases: Module

Conformalize a set of predictions. This is performed on hold-out data.

forward(p_scores)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

recalibrate(alpha=None)

Reset thresholds such that a new confidence limit is returned

Parameters:

alpha (float) – The power of the test / level of confidence interval

Returns:

qhat, the score threshold

Return type:

float

training: bool
dlsia.core.conformalize.conformalize_segmentation.does_not_have_label_in_set(conformal_set, label)

Checks if a conformal set does not contain a specific label

Parameters:

  • conformal_set (torch.Tensor, type bool (N,C, ...)) – The conformal set

  • label (torch.Tensor, type int) – the label of interest

Returns:

does it have this label

Return type:

torch.Tensor, type bool (N, …)

dlsia.core.conformalize.conformalize_segmentation.has_label_in_set(conformal_set, label)

Checks if a conformal set contains a specific label

Parameters:

  • conformal_set (torch.Tensor, type bool (N,C, ...)) – The conformal set

  • label (torch.Tensor, type int) – the label of interest

Returns:

does it have this label

Return type:

torch.Tensor, type bool (N, …)

dlsia.core.conformalize.conformalize_segmentation.index_it(scores, labs)

index the scores

Parameters:

  • scores – scores

  • labs – labels

Returns:

indexed scores

Return type:

dlsia.core.conformalize.conformalize_segmentation.set_size(conformal_set)

Compute the size of the conformal set, i.e. how many elements are in the set. The shape of the conformal_set tensor should be (N C …), we count the number of elements in channel 1. The conformal set is an indicator array, i.e. it contains True / False for each class.

Parameters:

conformal_set (torch.Tensor, type bool) – The conformal set

Returns:

The size of the conformal set

Return type:

(N, ….), type int

Module contents