Sample vp for Approximations, deprecate old ADVI #2027

ferrine · 2017-04-12T19:57:35Z

Following #2026

twiecki · 2017-04-13T15:44:10Z

pymc3/variational/advi.py

@@ -108,6 +108,9 @@ def advi(vars=None, start=None, model=None, n=5000, accurate_elbo=False,
        and Blei, D. M. (2016). Automatic Differentiation Variational
        Inference. arXiv preprint arXiv:1603.00788.
    """
+    import warnings
+    warnings.warn('Old ADVI interface is deprecated and be removed in future, use pm.ADVI instead',


and will be

ferrine · 2017-04-13T17:56:29Z

Seems like lots of refactoring will be needed for init_nuts

twiecki · 2017-04-14T15:08:45Z

pymc3/variational/advi.py

@@ -108,6 +108,10 @@ def advi(vars=None, start=None, model=None, n=5000, accurate_elbo=False,
        and Blei, D. M. (2016). Automatic Differentiation Variational
        Inference. arXiv preprint arXiv:1603.00788.
    """
+    import warnings
+    warnings.warn('Old ADVI interface is deprecated and will '
+                  'be removed in future, use pm.ADVI instead',


Shouldn't we tell people to use pm.fit() instead?

twiecki · 2017-04-14T15:10:52Z

pymc3/variational/callbacks.py

+        raise NotImplementedError
+
+
+class CheckLossConvergence(Callback):


Very elegant.

We should also add the classic ADVI convergence criterion: https://github.com/pymc-devs/pymc3/blob/master/pymc3/variational/advi.py#L180

twiecki · 2017-04-14T15:12:22Z

@ferrine This starts to look pretty good. Why is init_nuts trickier than what you already have here?

fonnesbeck · 2017-04-14T15:49:22Z

Do we want to do the sample_vp renaming here?

ferrine · 2017-04-14T17:50:23Z

@fonnesbeck The only option is to rename to sample_approx it is long and it's short name is weird

ferrine · 2017-04-14T17:51:06Z

@twiecki why do we need some strange tests for random states?

ferrine · 2017-04-14T18:01:07Z

@twiecki init nuts became a problem because of start point for fine-tuning advi from map. I did not support that case.
I also did not have cunstructor for inference from it's approx. It is done badly here. But I didn't find more pretty solution.
So refactoring adds more constructors and start point. I also want to try svgd init

fonnesbeck · 2017-04-14T18:04:04Z

@ferrine sample_approx isn't too long. I think being adequately descriptive and intuitive is more important than the number of characters, in general. I thought we wanted to generalize _vp away from just variational posteriors?

ferrine · 2017-04-14T18:15:53Z

@fonnesbeck good point. So method of Approximation will be renamed to just sample in order not to tell obvious things or misinform.

fonnesbeck · 2017-04-14T18:16:44Z

@ferrine that's a good idea, I agree.

ferrine · 2017-04-14T18:35:47Z

Comming back to the tests, they are failing. What for are these tests? https://github.com/pymc-devs/pymc3/blob/master/pymc3/tests/test_sampling.py#L37

twiecki · 2017-04-14T18:38:05Z

That test was added when we fixed a bug where every parallel sampler generated the same samplers because they all used the same seed.

ferrine · 2017-04-14T18:44:32Z

Hmm, interesting

ferrine · 2017-04-14T19:42:50Z

@twiecki I can't find where I use global numpy rng in my code :(

twiecki · 2017-04-14T19:43:19Z

@ferrine Does that test fail locally too?

ferrine · 2017-04-14T19:44:25Z

yes, for sanity check I used the following code:

s1 = np.random.get_state()
approx = pm.fit(
            n=10, method='advi', model=nnet,
            callbacks=[pm.callbacks.CheckLossConvergence(every=1, window_size=2)]
        )
start = approx.sample(draws=4)
cov = approx.cov.eval()
s2 = np.random.get_state()
(s1[0] == s2[0]), (s1[1] == s2[1]).all(), (s1[2] == s2[2]), (s1[3] == s2[3])
--------------
True, True, True, True

twiecki · 2017-04-14T19:45:38Z

Good thing we have that test.

@twiecki

Following discussion in #1953 CC @twiecki, @fonnesbeck, @aseyboldt, @jsalvatier, @taku-y, @springcoil

ferrine · 2017-04-20T21:16:05Z

I've added Inference module to docs, how can I check that it will be displayed?

fonnesbeck · 2017-04-20T21:18:37Z

Will there be a progress bar for sample_vp?

ferrine · 2017-04-20T21:20:22Z

Do you mean sample_approx? There is no need as it is vectorized and done only once for the whole sample size.

ferrine · 2017-04-21T18:00:07Z

What about merge?

ferrine · 2017-04-21T18:02:32Z

pymc3/variational/approximations.py

-        mapping {model_variable -> local_variable}
+    local_rv : dict[var->tuple]
+        Experimental for Empirical Distribution
+        mapping {model_variable -> local_variable (:math:`\\mu`, math:`\\rho`)}


fonnesbeck · 2017-04-22T17:59:41Z

Looks great, thanks!

twiecki · 2017-04-23T15:08:20Z

pymc3/sampling.py

-        v_params = pm.variational.advi(n=n_init, start=start,
-                                       random_seed=random_seed)
-        cov = np.power(model.dict_to_array(v_params.stds), 2)
+        approx = pm.MeanField(model=model, start=start)


I'm getting meaningful differences with this change on this model: http://twiecki.github.io/blog/2017/03/14/random-walk-deep-net/

It does not abort when converged and sampling from this initialization is very slow.

Do you see any way of improvement?

@twiecki I was suggested to try running minimum stopping criteria. Interesting to see how it will affect performance

twiecki reviewed Apr 13, 2017

View reviewed changes

twiecki reviewed Apr 14, 2017

View reviewed changes

ferrine added 11 commits April 14, 2017 22:07

add sample_vp, deprecate old ADVI

7b5d570

found typo

31e524d

fix docs

ea1cd82

add sample_vp to __init__

f952cdd

typo

b9d0a03

refactor init nuts, and needed stuff

35afe54

refactor sampling

5c4edd5

typo

5e12767

unused import

2b4442d

rename sample_approx

ac715bf

unused import

67fe071

allow seed kwarg

e9ce599

aseyboldt mentioned this pull request Apr 14, 2017

advi conversion detection for simple models #2031

Closed

ferrine added 4 commits April 15, 2017 01:40

add docks, additional callbacks

21c1615

change callback

d493caa

fix typo

d4fd914

check nan

ce70f4a

ferrine mentioned this pull request Apr 19, 2017

Optimization #1953

Closed

ferrine added 6 commits April 19, 2017 20:41

refactor callback

5e5756a

fix pylint, good catch

9c61bf2

Histogram -> Empirical (API change!)

daa9ee0

Following discussion in #1953 CC @twiecki, @fonnesbeck, @aseyboldt, @jsalvatier, @taku-y, @springcoil

refactor boilerplate Inference

8b23218

refactor callback

3e6311e

add progressbar supprot for init_nuts

c7e0f40

ferrine mentioned this pull request Apr 19, 2017

Stochastic Gradient Fisher Scoring #1977

Merged

ferrine added 4 commits April 20, 2017 22:45

launch tt_rng before return

5e750e7

Docs

f62fd25

add Inference to api reference

1d549f6

fix typo in doc

bc56865

make approximation docs more verbose

e808939

ferrine commented Apr 21, 2017

View reviewed changes

fix typos

8f99a2c

fonnesbeck merged commit 2e3caa1 into pymc-devs:master Apr 22, 2017

This was referenced Apr 22, 2017

sample_vp does not accept Histogram as argument #2026

Closed

Latest documentation is not displayed #2068

Closed

twiecki reviewed Apr 23, 2017

View reviewed changes

		raise NotImplementedError


		class CheckLossConvergence(Callback):

Uh oh!

Sample vp for Approximations, deprecate old ADVI #2027

Sample vp for Approximations, deprecate old ADVI #2027

Uh oh!

Conversation

ferrine commented Apr 12, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ferrine commented Apr 13, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

twiecki commented Apr 14, 2017

Uh oh!

fonnesbeck commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

fonnesbeck commented Apr 14, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

fonnesbeck commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

twiecki commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

twiecki commented Apr 14, 2017

Uh oh!

ferrine commented Apr 14, 2017

Uh oh!

twiecki commented Apr 14, 2017

Uh oh!

ferrine commented Apr 20, 2017

Uh oh!

fonnesbeck commented Apr 20, 2017

Uh oh!

ferrine commented Apr 20, 2017

Uh oh!

ferrine commented Apr 21, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

fonnesbeck commented Apr 22, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

fonnesbeck commented Apr 14, 2017 •

edited

Loading