Gamma collab #3890
Gamma collab #3890
Conversation
… pair program the rest
|
@jaketrookman sent me this link to the numpy version: https://github.com/numpy/numpy/blob/main/numpy/random/src/distributions/distributions.c#L220 |
| var b = kArg - 1/3; | ||
| var c = 1/sqrt(9 * b); | ||
| var X = 0.0; | ||
| var V = -1.0; |
There was a problem hiding this comment.
I think this is the bug. var V = -1.0; needs to go inside the while loop. Otherwise, it does V * V * V over and over again which eventually goes to infinity.
There was a problem hiding this comment.
On a first analysis, it looks like the V = -1.0 line was inserted so that the "while V <= 0" loop almost immediately below will execute at least once. But that could be done via a do-while loop, as is done in the python.
|
On further analysis, this looks like a faithful line by line copy of the python code, assuming that certain things correspond (e.g. bitgen_t in python to randomStream in chapel). The only difference I see is the V = -1 line that Amanda commented on above. On further further analysis, I did my best to test this as a standalone function, and so far I'm seeing no differences between the while-do and do-while versions. I was able to produce some anomalous seed behavior. The short description: my while-do and do-while versions almost always produce the same results given the same seeds. But sometimes they don't. Maybe when the holidays are over and we're all back in the office, I can show you what I mean. |
|
|
||
| } | ||
| else { | ||
| var b = kArg - 1/3; |
There was a problem hiding this comment.
1/3 needs to be 1/3.0 for correct floating point division. Same for any other computations in this function. This is
causing the distribution to be slightly off.
Work in progress of implementing Gamma function distribution into the random module
Need to look into setting seed, linter error, testing failure