netcal.scaling.LogisticCalibration

class netcal.scaling.LogisticCalibration(*args, temperature_only: bool = False, **kwargs)

On classification, apply the logistic calibration method aka Platt scaling to obtain a calibration mapping. This method is originally proposed by [1]. For the multiclass case, we use the Vector scaling proposed in [2]. On detection mode, this calibration method uses multiple independent normal distributions to obtain a calibration mapping by means of the confidence as well as additional features [3]. This calibration scheme assumes independence between all variables.

On detection, it is necessary to provide all data in input parameter X as an NumPy array of shape (n_samples, n_features), whereas the confidence must be the first feature given in the input array. The ground-truth samples y must be an array of shape (n_samples,) consisting of binary labels \(y \in \{0, 1\}\). Those labels indicate if the according sample has matched a ground truth box \(\text{m}=1\) or is a false prediction \(\text{m}=0\).

Mathematical background: For confidence calibration in classification tasks, a confidence mapping \(g\) is applied on top of a miscalibrated scoring classifier \(\hat{p} = h(x)\) to deliver a calibrated confidence score \(\hat{q} = g(h(x))\).

For detection calibration, we can also use the additional box regression output which we denote as \(\hat{r} \in [0, 1]^J\) with \(J\) as the number of dimensions used for the box encoding (e.g. \(J=4\) for x position, y position, width and height). Therefore, the calibration map is not only a function of the confidence score, but also of \(\hat{r}\). To define a general calibration map for binary problems, we use the logistic function and the combined input \(s = (\hat{p}, \hat{r})\) of size K by

\[g(s) = \frac{1}{1 + \exp(-z(s))} ,\]

According to [1], we can interpret the logit \(z\) as the logarithm of the posterior odds

\[z(s) = \log \frac{f(\text{m}=1 | s)}{f(\text{m}=0 | s)} \approx \log \frac{f(s | \text{m}=1)}{f(s | \text{m}=1)} = \ell r(s)\]

If we assume independence of all variables given in \(s\), we can use multiple univariate probability density distributions with the same variance to obtain a calibration mapping. Using this formulation, we can simply extend the scaling factor (from classification logistic calibration) to a scaling vector \(w \in \mathbb{R}^K\). However, instead of using the uncalibrated confidence estimate \(\hat{p}\), we use the logit of the network as part of \(s\) to be conform with the original formulation in [1] and [2]. Thus, the log-likelihood ratio can be expressed as

\[\ell r(s) = s^T w + c,\]

with bias \(c \in \mathbb{R}\). We utilize standard optimization methods to determine the calibration mapping \(g(s)\).

Capturing epistemic uncertainty of the calibration method is also able with this implementation [4].

Parameters:

• temperature_only (bool, default: False) – If True, use Temperature Scaling instead of Platt/Vector Scaling.

• method (str, default: "mle") – Method that is used to obtain a calibration mapping: - ‘mle’: Maximum likelihood estimate without uncertainty using a convex optimizer. - ‘momentum’: MLE estimate using Momentum optimizer for non-convex optimization. - ‘variational’: Variational Inference with uncertainty. - ‘mcmc’: Markov-Chain Monte-Carlo sampling with uncertainty.

• momentum_epochs (int, optional, default: 1000) – Number of epochs used by momentum optimizer.

• mcmc_steps (int, optional, default: 20) – Number of weight samples obtained by MCMC sampling.

• mcmc_chains (int, optional, default: 1) – Number of Markov-chains used in parallel for MCMC sampling (this will result in mcmc_steps * mcmc_chains samples).

• mcmc_warmup_steps (int, optional, default: 100) – Warmup steps used for MCMC sampling.

• vi_epochs (int, optional, default: 1000) – Number of epochs used for ELBO optimization.

• detection (bool, default: False) – If False, the input array ‘X’ is treated as multi-class confidence input (softmax) with shape (n_samples, [n_classes]). If True, the input array ‘X’ is treated as a box predictions with several box features (at least box confidence must be present) with shape (n_samples, [n_box_features]).

• independent_probabilities (bool, optional, default: False) – Boolean for multi class probabilities. If set to True, the probability estimates for each class are treated as independent of each other (sigmoid).

• use_cuda (str or bool, optional, default: False) – Specify if CUDA should be used. If str, you can also specify the device number like ‘cuda:0’, etc.

References

[1] (1,2,3)

Platt, John: “Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods.” Advances in large margin classifiers 10.3: 61-74, 1999 Get source online

[2] (1,2)

Chuan Guo, Geoff Pleiss, Yu Sun and Kilian Q. Weinberger: “On Calibration of Modern Neural Networks.” Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 2017. Get source online

[3]

Fabian Küppers, Jan Kronenberger, Amirhossein Shantia and Anselm Haselhoff: “Multivariate Confidence Calibration for Object Detection.” The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020.

[4]

Fabian Küppers, Jan Kronenberger, Jonas Schneider and Anselm Haselhoff: “Bayesian Confidence Calibration for Epistemic Uncertainty Modelling.” 2021 IEEE Intelligent Vehicles Symposium (IV), 2021

Methods

__init__(*args[, temperature_only])	Create an instance of LogisticCalibration.
clear()	Clear model parameters.
convex(data, y, tensorboard, log_dir)	Convex optimization to find the global optimum of current parameter search.
epsilon(dtype)	Get the smallest digit that is representable depending on the passed dtype (NumPy or PyTorch).
fit(X, y[, random_state, tensorboard, log_dir])	Build logitic calibration model either conventional with single MLE estimate or with Variational Inference (VI) or Markov-Chain Monte-Carlo (MCMC) algorithm to also obtain uncertainty estimates.
fit_transform(X[, y])	Fit to data, then transform it.
get_metadata_routing()	Get metadata routing of this object.
get_params([deep])	Get parameters for this estimator.
guide([X, y])	Variational substitution definition for each parameter.
load_model(filename)	Load model from saved torch dump.
mask()	Seek for all relevant weights whose values are negative.
mcmc(data, y, tensorboard, log_dir)	Perform Markov-Chain Monte-Carlo sampling on the (unknown) posterior.
model([X, y])	Definition of the log regression model.
momentum(data, y, tensorboard, log_dir)	Momentum optimization to find the global optimum of current parameter search.
prepare(X)	Preprocessing of input data before called at the beginning of the fit-function.
prior(dtype)	Prior definition of the weights used for log regression.
save_model(filename)	Save model instance as with torch's save function as this is safer for torch tensors.
set_fit_request(*[, log_dir, random_state, ...])	Request metadata passed to the fit method.
set_output(*[, transform])	Set output container.
set_params(**params)	Set the parameters of this estimator.
set_transform_request(*[, mean_estimate, ...])	Request metadata passed to the transform method.
to(device)	Set distribution parameters to the desired device in order to compute either on CPU or GPU.
transform(X[, num_samples, random_state, ...])	After model calibration, this function is used to get calibrated outputs of uncalibrated confidence estimates.
variational(data, y, tensorboard, log_dir)	Perform variational inference using the guide.