On 31/07/2019 00:49, Gilles Sadowski wrote:
Le mar. 30 juil. 2019 à 22:07, Alex Herbert <alex.d.herb...@gmail.com> a écrit :
On 30 Jul 2019, at 19:28, Gilles Sadowski <gillese...@gmail.com> wrote:
Hi.
Le mar. 30 juil. 2019 à 15:38, Alex Herbert <alex.d.herb...@gmail.com
<mailto:alex.d.herb...@gmail.com>> a écrit :
On 30/07/2019 10:56, Gilles Sadowski wrote:
Hello.
Le lun. 10 juin 2019 à 17:17, Alex Herbert <alex.d.herb...@gmail.com> a écrit :
On 10/06/2019 15:31, Gilles Sadowski wrote:
P.S. Thinking of releasing 1.3?
Not yet. I think there are a few outstanding items [...]
Anything missing?
- RNG-110: The PR for SharedSharedDiscrete/ContinuousSampler should have
a review [1]. I've left this while we finished GSoC phase 2 but it is ready.
I added factory methods for all samplers. For existing samplers this is
just for consistency. Some however use internal delegates and the
factory method can return the delegate directly which is an advantage.
One issue to look at is how I handled GaussianSampler and
LogNormalSampler. The samplers can only be shared state samplers if the
input NormalizedGaussianSampler is a shared state sampler. I handled
this with documentation. But this means a downstream user may be passed
a SharedStateContinuousSampler, use it as such and receive an exception
if it was created incorrectly.
Could the library create it "incorrectly”?
No. All the NormalizedGaussianSamplers in the library are suitable to pass to
the GaussianSampler and LogNormalSampler.
The alternative is two factory methods which must have different names
due to type erasure:
public static ContinuousSampler of(NormalizedGaussianSampler gaussian,
double scale, double shape);
public static
<T extends NormalizedGaussianSampler &
SharedStateSampler<ContinuousSampler>>
SharedStateContinuousSampler
ofSharedState(T normalized,
double mean,
double standardDeviation) {
Not nice, at first sight.
So the options are:
- As current but has the pitfall of throwing exceptions if you do create
a one with something that does not share state (i.e. not a sampler in
the library).
IIUC, it answers my question above.
I would not consider too much that the interfaces defined in our "client API"
module could be implemented by external codes.
Even within "Commons", we do not use other components' API…
OK. User beware documentation. If using only our library it will not be an
issue.
So I take it that you are fine with the PR?
Yes.
I didn't look at all the details but is that check
---CUT---
if (!(source.normalized instanceof SharedStateSampler<?>)) {
throw new UnsupportedOperationException("The underlying
sampler is not a SharedStateSampler");
}
---CUT---
necessary?
That was already in the codebase. I added it to support
SharedStateSampler in RNG-102.
The withUniformRandomProvider method is documented to:
@throws UnsupportedOperationException if the underlying sampler is not
a {@link SharedStateSampler}.
@throws ClassCastException if the underlying {@link SharedStateSampler}
does not return a {@link NormalizedGaussianSampler}.
So currently the two cases can be distinguished. If we drop the
source.normalized instanceof SharedStateSampler<?>
check then the next part will throw a ClassCastException if either case
is true.
So the question is really what exception should be raised in the rare
case that the operation cannot be performed (because the user created
with a NormalizedGaussianSampler not from the library). The
ClassCastException (which indicates the executing code did not check
what it can do) could be caught and rethrown as an
UnsupportedOperationException. Or the object could be tested with
instanceof before the cast. This would at least be consistent and show
the code knows what it is doing. So if a user calls:
GaussianSampler.withUniformRandomProvider(rng)
The only exception they have to handle is an
UnsupportedOperationException. An alternative would be an
IllegalStateException but that would indicate that the method would be
possible but not at this time. I think this would be better than a
ClassCastException.
My vote is to make UnsupportedOperationException the only exception to
be thrown. The ClassCastException to allow the cases to be separated is
unnecessary.
I can change this in master. It is only in GaussianSampler and
LogNormalSampler. There are already unit tests for the edge cases so
these just have to be updated to expect the appropriate documented
exception.
Just rereading it now and I’m not totally sold on the factory constructors
using .of(…). Here’s Joshua Bloch on the matter [1]:
—
Here are some common names for static factory methods:
• valueOf—Returns an instance that has, loosely speaking, the same
value as its parameters. Such static factories are effectively type-conversion
methods.
• of—A concise alternative to valueOf, popularized by EnumSet (Item
32).
• getInstance—Returns an instance that is described by the parameters
but cannot be said to have the same value. In the case of a singleton,
getInstance takes no parameters and returns the sole instance.
• newInstance—Like getInstance, except that newInstance guarantees
that each instance returned is distinct from all others.
• getType—Like getInstance, but used when the factory method is in a
different class. Type indicates the type of object returned by the factory
method.
• newType—Like newInstance, but used when the factory method is in a
different class. Type indicates the type of object returned by the factory
method.
—
The ’newType’ may be appropriate as the factory method is returning an instance
of an interface not defined in the class. This is required for some
implementations such as the PoissonSampler which returns either a SmallMean or
LargeMeanPoissonSampler. So the Bloch naming could be ’newSampler’. But then we
have over verbose code using ‘Sampler' twice:
new PoissonSampler(rng, mean)
PoissonSampler.of(rng, mean)
PoissonSampler.newSampler(rng, mean)
IMO the later is too verbose. If we state that the sampler is entirely defined by
the input arguments to ‘of’ then it does satisfy "has, loosely speaking, the
same value as its parameters”.
So ‘of’ is OK. The only other words I like are ‘from’ or ‘with’:
PoissonSampler.of(rng, mean)
PoissonSampler.from(rng, mean)
PoissonSampler.with(rng, mean)
"with" seems the most appropriate but has the disadvantage of not
being in Bloch's list.
I would then stick with "of" because "from" looks like the arguments
will be transformed
(and it also is not in the list...).
Gilles
WDYT?
[1] http://www.informit.com/articles/article.aspx?p=1216151
<http://www.informit.com/articles/article.aspx?p=1216151>
- Another factory method to explicitly create a SharedStateSampler using
a normalised Gaussian SharedStateSampler.
A few things that are 90% done:
- RNG-85: MiddleSquareWeylSequence generator
This is simple code and now the modifications have been made to the
ProviderBuilder it is possible to pass in a good quality increment for
the Weyl sequence. I have code to build the increment that can be added
to the SeedFactory. I did this months ago so will have to find it and
create the PR.
- RNG-95: DiscreteUniformSampler
I now have a reference for the alternative algorithm for choosing int
values from an interval. The code is done but should go after RNG-110 as
the code uses 5 internal delegates for different algorithms. This would
be optimised by the changes in RNG-110.
- RNG-109: DiscreteProbabilityCollectionSampler to use an internal
DiscreteSampler
I have to create a benchmark to compare the AliasMethodSampler against
the GuideTableSampler to see which is more suitable for a generic
probability distribution. This should not take long.
- RNG-94: RotateRotateMultiplyXorMultiplyXor
Simple code that is based on the same idea of using an output hash
function on a Weyl sequence like SplitMix. It is slightly slower but the
hash function is better and more robust to low complexity increments. So
we can add it using a seeded increment for the Weyl sequence. This would
take a day to add the two hash function variants.
Everything ready is fine to add before the next release, but equally
fine to add after it (and do another release in 2 months if wanted)
given the host of new features already implemented. :-)
I’ll put what I have together in the coming week or so.
Best,
Gilles
Maybe for later:
- RNG-90: Improve nextInt(int)
This could use the same algorithm as RNG-95. I have not done the testing
yet. It also can be done for nextLong(long) which requires a 64-bit
product multiplication to be computed as a 128-bit result. I have code
for this but no performance tests.
Not done but...
The PCG family has extended generators: K-dimensionally equidistributed
or Cryptographic. These have a much larger period and the
equidistributed ones can be Jumpable.
[1] https://github.com/apache/commons-rng/pull/58
Regards,
Gilles
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
For additional commands, e-mail: dev-h...@commons.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
For additional commands, e-mail: dev-h...@commons.apache.org
