View source on GitHub

A transformation of samples of the outer BayesianModel.

Specifically, E_{x~p}[f(x)] for a model p and transformation f. The model p is implicit, in that the SampleTransformation appears in the sample_transformations field of that BayesianModel. The f is given as fn so that candidate samples may be passed through it. The fn may close over the parameters of p, and the ground_truth_mean will presumably depend on p implicitly via some sampling process.

If the ground_truth_mean is estimated by sampling, then ground_truth_standard_deviation and ground_truth_mean_standard_error are related using the standard formula:

SEM = SD / sqrt(N)

where N is the number of samples. ground_truth_standard_deviation describes the distribution of f(x), while ground_truth_mean_standard_error desribes how accurately we know ground_truth_mean.


An identity fn for a vector-valued target would look like:

fn = lambda x: x

fn Function that takes samples from the target and returns a (nest of) Tensor. The returned Tensor must retain the leading non-event dimensions.
pretty_name Human readable name, suitable for a table in a paper.
ground_truth_mean Ground truth value of this expectation. Can be None if not available. Default: None.
ground_truth_mean_standard_error Standard error of the ground truth mean. Can be None if not available. Default: None.
ground_truth_standard_deviation Standard deviation of samples transformed by fn. Can be None if not available. Default: None.
ground_truth_standard_deviation_standard_error Standard error of the ground truth standard deviation. Can be None if not available. Default: None.
dtype Possibly nested dtype of the output of fn. Default: tf.float32.



View source

Returns fn(value).