multinomial about mir-random

[jmh530]

It's like you port it more-or-less straight from GSL. The only line I'm concerned about is auto rng = Random(25923); I would think you would instead have something like alias rng = rne; but I don't use mir.random that much as of now...so I don't know what is the recommended approach. As you say, it would definitely need to be modified to be part of mir.random's API.

…

On Tue, Oct 1, 2019 at 9:47 AM Flávio Codeço Coelho < ***@***.***> wrote: Ok, I'll be glad to contribute. I have started an implementation ported straight from GSL <https://github.com/ampl/gsl/blob/master/randist/multinomial.c>. the code is here: https://gist.github.com/run-dlang/edc536446341a957256d1bfeeefaf1da However, there is still some bug in it, if someone could have a look I'd *greatly* appreciate it. After we get this working, It'll still need to be modified to conform with mir.random's API, but I believe that as is it is easier to test and make sure is working properly. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#117?email_source=notifications&email_token=ADBFNSVNOGHRWGWSZ5PMEKDQMNIHJA5CNFSM4I3UG7S2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEABKJ7A#issuecomment-537044220>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADBFNSXWSFB4STD6XKMAXJ3QMNIHJANCNFSM4I3UG7SQ> .

[9il]

We don't, but we can review and accept a pull request with this random variable.

[jmh530]

This shouldn't be that hard to implement on your own given BinomialVariable. According to the r function's documentation [1], they calculate it as:

The rmultinom() algorithm draws binomials X[j] from Bin(n[j], P[j]) sequentially, where n[1] = N (N := size), P[1] = p[1] (p is prob scaled to sum 1), and for j ≥ 2, recursively, n[j] = N - sum(k=1, …, j-1) X[k] and P[j] = p[j] / (1 - sum(p[1:(j-1)])).

[1] https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Multinom.html

[fccoelho]

Ok,

I'll be glad to contribute. I have started an implementation ported straight from GSL.
the code is here: https://gist.github.com/run-dlang/edc536446341a957256d1bfeeefaf1da
However, there is still some bug in it, if someone could have a look I'd greatly appreciate it.
After we get this working, It'll still need to be modified to conform with mir.random's API, but I believe that as is it is easier to test and make sure is working properly.

[jmh530]

One other change it needs: change [1.0/6, 2.0/6, 3/6] to [1.0/6, 2.0/6, 3.0/6] On Wed, Oct 2, 2019 at 10:31 PM John Hall <[email protected]> wrote:

…

It's like you port it more-or-less straight from GSL. The only line I'm concerned about is auto rng = Random(25923); I would think you would instead have something like alias rng = rne; but I don't use mir.random that much as of now...so I don't know what is the recommended approach. As you say, it would definitely need to be modified to be part of mir.random's API. On Tue, Oct 1, 2019 at 9:47 AM Flávio Codeço Coelho < ***@***.***> wrote: > Ok, > > I'll be glad to contribute. I have started an implementation ported > straight from GSL > <https://github.com/ampl/gsl/blob/master/randist/multinomial.c>. > the code is here: > https://gist.github.com/run-dlang/edc536446341a957256d1bfeeefaf1da > However, there is still some bug in it, if someone could have a look I'd > *greatly* appreciate it. > After we get this working, It'll still need to be modified to conform > with mir.random's API, but I believe that as is it is easier to test and > make sure is working properly. > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub > <#117?email_source=notifications&email_token=ADBFNSVNOGHRWGWSZ5PMEKDQMNIHJA5CNFSM4I3UG7S2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEABKJ7A#issuecomment-537044220>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ADBFNSXWSFB4STD6XKMAXJ3QMNIHJANCNFSM4I3UG7SQ> > . >

[fccoelho]

Thanks @jmh530, for spotting the errors! It seems to be working fine now, I'll come up with some unit tests now to make sure it is probabilistically correct, i.e., the sampled values match theoretical expectations.

Meanwhile, should I use this as a code template, for the pull request?
I am new to D so I'll have to take some time to digest some of the specifics of the mir-random code.

[jmh530]

@fccoelho I would definitely take a look at variable.d to get a sense of how each of these are written, but ultimately it should be in ndvariable.d and follow what is going on there.

[fccoelho]

Thanks, it makes sense.

[fccoelho]

I have got some preliminary implementation done, but as it stands it is not compiling because I can't figure out how templates work to deduce argument types.

I could use some help here.

[jmh530]

The first thing I noticed is that you used T instead of ulong in the signature here for N (making it a U that is constrained to be able to convert to ulong might be better) MultinomialVariable!T multinomialVar(T)(in T N, in T[] probs) However, it still doesn't compile, but that's a relatively easy fix in the unittest to make sure the call to rv is double[3] result; rv(rne, result[]); I did notice that the MultivariateNormalVariable opCall returns void and yours doesn't. It also takes into account mir slices.

…

On Thu, Oct 3, 2019 at 3:54 PM Flávio Codeço Coelho < ***@***.***> wrote: I have got some preliminary implementation <https://github.com/fccoelho/mir-random/blob/f01e87c8b5966527aeddc49a76517502de7cfc43/source/mir/random/ndvariable.d#L283> done, but as it stands it is not compiling because I can't figure out how templates work to deduce argument types. I could use some help here. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#117?email_source=notifications&email_token=ADBFNSU4SB2B5HHJQVJDTR3QMZEYFA5CNFSM4I3UG7S2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEAJMUKY#issuecomment-538102315>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADBFNSTIGODCN6KJ3FJYALLQMZEYFANCNFSM4I3UG7SQ> .

[fccoelho]

Ok, I got it to work, thanks again @jmh530 .
For some reason I can get the test assert(x[0]+x[1]+x[2] == s); to work any Ideas of how I can get this simple line to run? It seems to be related to the way result is declared in the opCall (with scope)

[jmh530]

I can't say for sure, but typically you need to use something like approxEqual when dealing with equality checks for floating point data. https://dlang.org/library/std/math/approx_equal.html

…

On Fri, Oct 4, 2019 at 8:55 AM Flávio Codeço Coelho < ***@***.***> wrote: Ok, I got it to work, thanks again @jmh530 <https://github.com/jmh530> . For some reason I can get the test assert(x[0]+x[1]+x[2] == s); to work any Ideas of how I can get this simple line to run? It seems to be related to the way result is declared in the opCall (with scope) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#117?email_source=notifications&email_token=ADBFNSRQKIYHK6GHQLFOVGDQM44K3A5CNFSM4I3UG7S2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEALRSHA#issuecomment-538384668>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADBFNSS2DUHUJ7QITQS2GTDQM44K3ANCNFSM4I3UG7SQ> .

[fccoelho]

I should have added more information: it is not that the test fails. it fails to compile
here is the test as it currently stands:

nothrow @safe version(mir_random_test) unittest
{
    Random* gen = threadLocalPtr!Random;
    ulong s = 1000;
    auto rv = MultinomialVariable!(ulong,double)(s, [1.0, 5.7, 0.3]);
    ulong[3] x;
    rv(gen,x);
    assert(x[0]+x[1]+x[2] == s);
    assert(1==1);
}

when I try to dub test the package this is what I get:

Running ./mir-random-test-default 
core.exception.AssertError@source/mir/random/ndvariable.d(397): unittest failure
----------------
??:? _d_unittestp [0x559ccc585601]
source/mir/random/ndvariable.d:397 nothrow @safe void mir.random.ndvariable.__unittest_L390_C40() [0x559ccc576b35]
??:? void mir.random.ndvariable.__modtest() [0x559ccc578218]
??:? int core.runtime.runModuleUnitTests().__foreachbody2(object.ModuleInfo*) [0x559ccc5a0d23]
??:? int object.ModuleInfo.opApply(scope int delegate(object.ModuleInfo*)).__lambda2(immutable(object.ModuleInfo*)) [0x559ccc583d7e]
??:? int rt.minfo.moduleinfos_apply(scope int delegate(immutable(object.ModuleInfo*))).__foreachbody2(ref rt.sections_elf_shared.DSO) [0x559ccc58b6d5]
??:? int rt.sections_elf_shared.DSO.opApply(scope int delegate(ref rt.sections_elf_shared.DSO)) [0x559ccc58b868]
??:? int rt.minfo.moduleinfos_apply(scope int delegate(immutable(object.ModuleInfo*))) [0x559ccc58b661]
??:? int object.ModuleInfo.opApply(scope int delegate(object.ModuleInfo*)) [0x559ccc583d55]
??:? runModuleUnitTests [0x559ccc5a0b01]
??:? void rt.dmain2._d_run_main(int, char**, extern (C) int function(char[][])*).runAll() [0x559ccc5875e0]
??:? void rt.dmain2._d_run_main(int, char**, extern (C) int function(char[][])*).tryExec(scope void delegate()) [0x559ccc58756b]
??:? _d_run_main [0x559ccc5874d1]
??:? main [0x559ccc4f9015]
??:? __libc_start_main [0x7f263687a09a]
1/16 unittests FAILED
Program exited with code 1

Line 397 is exactly that assert.

[jmh530]

Ah, so it's not floating point and thus approxEqual shouldn't matter. I don't know for sure right now and am working and busy later tonight so I may not be able to look at it more until tomorrow. I would suggest at least taking a look at what the result of x actually is before doing that assert. That's just basic debugging of these sorts of things.

…

On Fri, Oct 4, 2019 at 9:26 AM Flávio Codeço Coelho < ***@***.***> wrote: I should have added more information: it is not that the test fails. it fails to compile here is the test as it currently stands: nothrow @safe version(mir_random_test) unittest { Random* gen = threadLocalPtr!Random; ulong s = 1000; auto rv = MultinomialVariable!(ulong,double)(s, [1.0, 5.7, 0.3]); ulong[3] x; rv(gen,x); assert(x[0]+x[1]+x[2] == s); assert(1==1); } when I try to dub test the package this is what I get: Running ./mir-random-test-default ***@***.***/mir/random/ndvariable.d(397): unittest failure---------------- ??:? _d_unittestp [0x559ccc585601] source/mir/random/ndvariable.d:397 nothrow @safe void mir.random.ndvariable.__unittest_L390_C40() [0x559ccc576b35] ??:? void mir.random.ndvariable.__modtest() [0x559ccc578218] ??:? int core.runtime.runModuleUnitTests().__foreachbody2(object.ModuleInfo*) [0x559ccc5a0d23] ??:? int object.ModuleInfo.opApply(scope int delegate(object.ModuleInfo*)).__lambda2(immutable(object.ModuleInfo*)) [0x559ccc583d7e] ??:? int rt.minfo.moduleinfos_apply(scope int delegate(immutable(object.ModuleInfo*))).__foreachbody2(ref rt.sections_elf_shared.DSO) [0x559ccc58b6d5] ??:? int rt.sections_elf_shared.DSO.opApply(scope int delegate(ref rt.sections_elf_shared.DSO)) [0x559ccc58b868] ??:? int rt.minfo.moduleinfos_apply(scope int delegate(immutable(object.ModuleInfo*))) [0x559ccc58b661] ??:? int object.ModuleInfo.opApply(scope int delegate(object.ModuleInfo*)) [0x559ccc583d55] ??:? runModuleUnitTests [0x559ccc5a0b01] ??:? void rt.dmain2._d_run_main(int, char**, extern (C) int function(char[][])*).runAll() [0x559ccc5875e0] ??:? void rt.dmain2._d_run_main(int, char**, extern (C) int function(char[][])*).tryExec(scope void delegate()) [0x559ccc58756b] ??:? _d_run_main [0x559ccc5874d1] ??:? main [0x559ccc4f9015] ??:? __libc_start_main [0x7f263687a09a]1/16 unittests FAILED Program exited with code 1 Line 397 is exactly that assert. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#117?email_source=notifications&email_token=ADBFNSXMW73QB7JJB76U6LLQM5AADA5CNFSM4I3UG7S2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEALUJJQ#issuecomment-538395814>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADBFNSTYXGVCS5JDXZ3ALELQM5AADANCNFSM4I3UG7SQ> .

[fccoelho]

Fixed everything apparently. I've just created a PR.

[9il]

The PR should pass before review

[9il]

Ah, ok, we need to update CI first

[fccoelho]

From what I could find from the Travis error log, the tests are failing when building for x86 architecture:

 dub test -ax86
Generating test runner configuration 'mir-random-test-default' for 'default' (library).
Excluding package.d file from test due to https://issues.dlang.org/show_bug.cgi?id=11847
Excluding package.d file from test due to https://issues.dlang.org/show_bug.cgi?id=11847
Excluding package.d file from test due to https://issues.dlang.org/show_bug.cgi?id=11847
Performing "unittest" build using dmd for x86.
mir-core 1.0.2: target for configuration "library" is up to date.
mir-algorithm 3.5.6: target for configuration "default" is up to date.
mir-linux-kernel 1.0.1: target for configuration "library" is up to date.
mir-random 2.2.6+commit.5.gee30a11: building configuration "mir-random-test-default"...
source/mir/random/ndvariable.d(344,38): Error: template mir.random.variable.binomialVar cannot deduce function from argument types !()(ulong, double), candidates are:
source/mir/random/variable.d(1750,20):        mir.random.variable.binomialVar(T)(size_t r, in T p) if (isFloatingPoint!T)
source/mir/random/ndvariable.d(388,11): Error: template instance `mir.random.ndvariable.MultinomialVariable!(ulong, double).MultinomialVariable.opCall!(MersenneTwisterEngine!(uint, 32u, 624u, 397u, 31u, 2567483615u, 11u, 4294967295u, 7u, 2636928640u, 15u, 4022730752u, 18u, 1812433253u))` error instantiating
source/mir/random/ndvariable.d(400,11): Error: template instance `mir.random.ndvariable.MultinomialVariable!(ulong, double).MultinomialVariable.opCall!(MersenneTwisterEngine!(uint, 32u, 624u, 397u, 31u, 2567483615u, 11u, 4294967295u, 7u, 2636928640u, 15u, 4022730752u, 18u, 1812433253u))` error instantiating
dmd failed with exit code 1.

I can't decifer what is going on here, if anyone can point me to what this error is about, I can make a quick commit, fixing this issue.

[9il]

ulong is always 64 bit. size_t is 32 bit on x86 and can't accept ulong argument.

[fccoelho]

I suspected as much. I have changed the N argument in the MultinomialVariable class frm ulong to uint and now I am getting an issue building for x86_64:

dub test
Generating test runner configuration 'mir-random-test-default' for 'default' (library).
Excluding package.d file from test due to https://issues.dlang.org/show_bug.cgi?id=11847
Excluding package.d file from test due to https://issues.dlang.org/show_bug.cgi?id=11847
Excluding package.d file from test due to https://issues.dlang.org/show_bug.cgi?id=11847
Performing "unittest" build using dmd for x86_64.
mir-core 1.0.2: target for configuration "library" is up to date.
mir-algorithm 3.5.6: target for configuration "default" is up to date.
mir-linux-kernel 1.0.1: target for configuration "library" is up to date.
mir-random 2.2.6+commit.5.gee30a11: building configuration "mir-random-test-default"...
source/mir/random/ndvariable.d(332,17): Error: no type for typeid(threadLocal)
source/mir/random/ndvariable.d(344,26): Error: cannot implicitly convert expression rv.opCall(threadLocal()) of type ulong to uint
source/mir/random/ndvariable.d(387,11): Error: template instance `mir.random.ndvariable.MultinomialVariable!(uint, double).MultinomialVariable.opCall!(MersenneTwisterEngine!(ulong, 64LU, 312LU, 156LU, 31LU, 13043109905998158313LU, 29LU, 6148914691236517205LU, 17LU, 8202884508482404352LU, 37LU, 18444473444759240704LU, 43LU, 6364136223846793005LU))` error instantiating
source/mir/random/ndvariable.d(399,11): Error: template instance `mir.random.ndvariable.MultinomialVariable!(uint, double).MultinomialVariable.opCall!(MersenneTwisterEngine!(ulong, 64LU, 312LU, 156LU, 31LU, 13043109905998158313LU, 29LU, 6148914691236517205LU, 17LU, 8202884508482404352LU, 37LU, 18444473444759240704LU, 43LU, 6364136223846793005LU))` error instantiating
dmd failed with exit code 1

It seems that the MersennTwisterEngine is expecting an ulong !

this seems to be a catch-22 type of situation..... If I set N to ulong I can't run it on x86 if I set it to uint, I cant run on x86_64

[jmh530]

@fccoelho Where the ulongs or uints are causing errors, try changing them to size_t.

[fccoelho]

Thanks, @jmh530. That solved it.

Now the PR is passing the Appveyor test but failing on Travis, but the log shows that the error is not about the D code but the build script.

[9il]

released in v2.2.8