Expand the testing coverage using R as the reference #414
Conversation
|
Good call |
- Erlang - FDist - Gamma - Normal - NormalCanon - TDist - Uniform
|
Thanks for all of your work here, @lindahua, this is great. Glad to see you around these parts again. 🙂 |
|
This PR is ready, and passed all tests in my machine. Now almost all distributions have been covered by this R-based reference framework, including many distributions that were not available in scipy, e.g. Generalized Extreme Value, Beta Binomial, Noncentral Beta/Chi squared/F/T, Noncentral Hypergeometric, Normal Inverse Gaussian, and Von Mises, etc. Also, with the new test cases, several bugs have been found and fixed. A readme describing this testing framework was provided. |
lindahua
changed the title
| end | ||
|
|
||
| kurtosis{T<:Real}(d::TriangularDist{T}) = T(-3)/5 | ||
|
|
||
| entropy(d::TriangularDist) = 1//2 + log((d.b - d.a) / 2) | ||
| entropy(d::TriangularDist) = 0.5 + log((d.b - d.a) / 2) |
There was a problem hiding this comment.
Shouldn't use Float64 literal here to avoid promoting e.g. Float32 input. Maybe one(d.b)/2.
| @@ -82,12 +82,12 @@ end | |||
|
|
|||
| function skewness(d::TriangularDist) | |||
| (a, b, c) = params(d) | |||
| sqrt2 * (a + b - 2c) * (2a - b - c) * (a - 2b + c) / (5 * _pretvar(a, b, c)^3//2) | |||
| sqrt2 * (a + b - 2c) * (2a - b - c) * (a - 2b + c) / (5 * _pretvar(a, b, c)^1.5) | |||
|
@andreasnoack Fix per your comments. |
|
The failures on master are because of JuliaLang/julia#20344 |
Currently, we use Python's Scipy package as a reference to verify the correctness of our implementation of distributions.
However, Scipy only provides a limited coverage of what we currently have. This PR aims to expand the coverage of our testing by switching to using R as our reference. R, together with the packages in CRAN, can cover most (if not all) of the distributions.