add stein methods

[ferrine]

More and more papers use stein methods. Here I implement convenient interface to it

[twiecki]

what's ASVGD?

[ferrine]

Amortized Stein Variational Gradient Descent
Can be found here. It is a novel black box algorithm for wild inference that doesn't require q(z)

[twiecki]

I think we removed all * imports in other parts of the code-base, shouldn't reintroduce it here. cc @ColCarroll

[ferrine]

That's ready for review

[twiecki]

I assigned @taku-y if he has time.

[ferrine]

I think I'll add a note about behavior of sampler in high dimensions. After I see no problem for merge
CC @taku-y

[ferrine]

Important Note:
For BEST model with FullRank approx I found an exact solution with large number of particles(1000). With low number of them variance was underestimated. So I'm not sure if it really works as proposed in the paper(theoretically). CC @lewisKit

Blue is ASVGD on FullRank approx, green is FullRankADVI

[taku-y]

Here is the log of Travis on TestSVGD.test_optimizer_minibatch_with_generator:

>       np.testing.assert_allclose(np.std(trace['mu']), np.sqrt(1. / d), rtol=0.4)
E       AssertionError: 
E       Not equal to tolerance rtol=0.4, atol=0
E       
E       (mismatch 100.0%)
E        x: array(0.13835382035008825)
E        y: array(0.09476178269858855)

I'm not sure if this error rate is acceptable.

[ferrine]

I'm more concerned about poor ASVGD performance on simple problems. I'l increase number of iterations for these tests and lower relative tolerance

[ferrine]

@taku-y I've fixed tests, they look better now. (passing locally)

[taku-y]

ASVGD results are comparable on the BEST example and the code looks good to me. I think its ready to merge.

[twiecki]

Awesome! Should also add an example with short description.

[ferrine]

Sure, coming soon in my blog

[ferrine]

I'll point on the problems that can be met using asvgd.

[springcoil]

Excellent work all

…

On 22 May 2017 10:17 pm, "Maxim Kochurov" ***@***.***> wrote: I'll point on the problems that can be met using asvgd. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2183 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA8DiDEt2OQlPKqJn1ww9fQFh5yLdwONks5r8ft5gaJpZM4NaE39> .