Added support for sampling from a Beta distribution. #21829

eaplatanios · 2019-01-13T16:18:55Z

This PR adds support for sampling from a Beta distribution. Are the Greek letters acceptable in the public API? If not I can replace them.

As a side note, I believe that we should ideally have a Distribution protocol that let's one sample from a distribution, but also compute other quantities like expectation, pdf, cdf, etc. For that reason, I believe a couple of interesting TODOs with respect to the Random module would be:

Generalize distributions over both Swift floating point types, but also TensorFlow floating point types (to allow for support of things like TensorFlow Probability).
Create a Distribution protocol that requires some more things for each distribution, such as computing expectation, pdf, cdf, etc.

rxwei · 2019-01-13T16:22:34Z

Are the Greek letters acceptable in the public API? If not I can replace them.

Not yet, probably in the (distant) future. Replacing them sounds great.

stdlib/public/TensorFlow/Random.swift

rxwei · 2019-01-13T16:27:07Z

Thanks Anthony! Will take a closer look later today.

rxwei · 2019-01-14T05:45:37Z

stdlib/public/TensorFlow/Random.swift

+          w = a * exp(v)
+          break
+        }
+


How about removing this empty line?

rxwei · 2019-01-14T05:47:14Z

stdlib/public/TensorFlow/Random.swift

+  /// algorithm, when both alpha and beta are greater than 1.
+  ///
+  /// - Parameters:
+  ///     - alpha: First Beta distribution shape parameter.


Align - with "Parameters".

Suggested change

/// - alpha: First Beta distribution shape parameter.

/// - alpha: First Beta distribution shape parameter.

rxwei · 2019-01-14T12:07:26Z

Our CI is down. We'll have to wait for it.

rxwei · 2019-01-17T00:58:14Z

@eaplatanios I think we also need a protocol that unifies all distributions so that Tensor random initializers can take a generic distribution and just use that.

eaplatanios · 2019-01-17T01:36:48Z

@rxwei I agree. I'll look into proposing one such protocol. I could do it as part of this PR but it will probably have to wait until after the ICML deadline (January 23rd) because I'm a bit busy until then.

It would also be nice to allow distributions to be defined over tensors (e.g., passing in a TensorFlow tensor as the mean of a distribution and returning TensorFlow tensors as samples).

rxwei · 2019-01-17T02:45:30Z