input.noise package

Submodules

input.noise.noise_process module

Created on 28 de mar de 2019

@author: klaus

class input.noise.noise_process.LabelNoiseProcess(**kwargs)

Bases: object

A noise process, which is used to corrupt the original, clean labels.

Currently, the following parameters are expected:

  • type : Specifies the type of noise. Currently, this may be

    1. "NCAR": Noisy Completely at Random. The occurrence of error is independent of other random variables, including the true labels. The chance of a label being flipped to a given, different class is :math:\frac{p}{C-1}. Requires specification of parameter noise_corruption, which determines \(p\).

  • seed : Specifies the seed for reproducibility purposes.

  • deterministic : If True, the distribution of actual flips tries to reflect the scaled matrix of transition probabilities. This means that, while the chosen labels might differ from run to run, the number of flips is essentially fixed, we have a fixed quantity of ‘corrupted labels’.

__init__(**kwargs)

Constructor for LabelNoiseProcess.

Parameters

**kwargs – Key-value pairs with the configuration options.

Raises

KeyError – If one of the required keys is not found.

corrupt(Y, labeledIndexes, hook=None)

Corrupts a set of clean labels, obtaining the corrupted labels.

Parameters

  • Y (NDArray[float].shape[N,C]) – Target belief matrix, whose rows are one-hot selecting the correct label.

  • labeledIndexes (NDArray[bool].shape[N]) – A boolean array, indicating which instances are to be interpreted as labeled.

  • hook (GSSLHook) – Optional. A hook to be called before/during/after the noise process

Returns

Corrupted version of the target belief matrix

Return type

(NDArray[float].shape[N,C])

input.noise.noise_utils module

noise_utils.py

Module containing utilities related to noise.

input.noise.noise_utils.apply_noise(Y, labeledIndexes, A, seed=None, deterministic=True)

Corrupts a set percentage of initial labels with noise.

Parameters

  • Y ([NDArray[int].shape[N,C]) – Matrix encoding initial beliefs.

  • A ([NDArray[int].shape[C,C]) – Transition probabilities between each class.

  • labeledIndexes (NDArray[bool].shape[N]) – determines which indices are to be considered as labeled.

  • seed (float) – Optional. Used to reproduce results.

Returns

Belief matrix after corruption.

Return type

NDArray[int].shape[N,C]

input.noise.noise_utils.transition_count_mat(Y, A)

Obtains a transition count matrix for uniform noise, indicating how many instances should have their label flipped and to which class.

Specifically, this returns a matrix M such that \(M[i,j]\) is the number of instances of i-th class to be swapped to the j-th class.

Parameters

  • Y ([NDArray[int].shape[N,C]) – Matrix encoding initial beliefs.

  • A ([NDArray[int].shape[C,C]) – Transition probabilities between each class.

Returns

the transition count matrix.

Return type

[NDArray[int].shape[C,C]

Raises

ValueError – if p is an invalid percentage.

input.noise.noise_utils.uniform_noise_transition_prob_mat(Y, p)

Obtains the transition probabilities between each class for uniform label noise.

Parameters

  • Y ([NDArray[int].shape[N,C]) – Matrix encoding initial beliefs.

  • p (float) – The percentage of labels to be flipped

Returns

the transition probability matrix.

Return type

[NDArray[int].shape[C,C]

Raises

ValueError – if p is an invalid percentage.

Module contents