tfp.substrates.numpy.stats.expected_calibration_error

Compute the Expected Calibration Error (ECE).

View aliases

Main aliases

tfp.experimental.substrates.numpy.stats.expected_calibration_error

tfp.substrates.numpy.stats.expected_calibration_error(
    num_bins, logits=None, labels_true=None, labels_predicted=None, name=None
)

This method implements equation (3) in [1]. In this equation the probability of the decided label being correct is used to estimate the calibration property of the predictor.

Note: a trade-off exist between using a small number of num_bins and the estimation reliability of the ECE. In particular, this method may produce unreliable ECE estimates in case there are few samples available in some bins. As an alternative to this method, consider also using bayesian_expected_calibration_error.

References

[1]: Chuan Guo, Geoff Pleiss, Yu Sun, Kilian Q. Weinberger, On Calibration of Modern Neural Networks. Proceedings of the 34th International Conference on Machine Learning (ICML 2017). arXiv:1706.04599 https://arxiv.org/pdf/1706.04599.pdf

Args
num_bins	int, number of probability bins, e.g. 10.
logits	Tensor, (n,nlabels), with logits for n instances and nlabels.
labels_true	Tensor, (n,), with tf.int32 or tf.int64 elements containing ground truth class labels in the range [0,nlabels].
labels_predicted	Tensor, (n,), with tf.int32 or tf.int64 elements containing decisions of the predictive system. If None, we will use the argmax decision using the logits.
name	Python str name prefixed to Ops created by this function.

Returns
ece	Tensor, scalar, tf.float32.