Convert zero-inflated distributions to Mixture subclasses #2231

fonnesbeck · 2017-05-27T19:47:43Z

Added the ZIB distribution to discrete.py, along with tests.

* Make spacing more compliant to PEP8 * Raise NotImplementedError from SplineWrapper grad operation Fixes #2209 * Instantiate SplineWrapper recursively to simplify the architecture * Create spline derivatives lazily * Move grad_op to a separate property * Add tests for SplineWrapper * Fix style issues

Update create testenv

* add draft * add draft * move and rename * better docstring * added tests

WIP: replace Hessian scaling guessing by an identity matrix

remaining fix in #2110

…tity" This reverts commit eb6042f, reversing changes made to c3afa00.

junpenglao · 2017-05-29T13:20:46Z

pymc3/tests/test_distributions.py

@@ -591,6 +591,10 @@ def test_zeroinflatednegativebinomial(self):
        self.checkd(ZeroInflatedNegativeBinomial, Nat,
                    {'mu': Rplusbig, 'alpha': Rplusbig, 'psi': Unit})

+    def test_zeroinflatedbinomial(self):
+        self.checkd(ZeroInflatedBinomial, Nat,


You forgot to import ZeroInflatedBinomial in the test_distributions.py

twiecki · 2017-05-29T13:23:01Z

Too bad we still haven't figured out how to do this with the Mixture class.

scaling for NUTS using 'advi_map'

fonnesbeck · 2017-05-29T14:07:17Z

Let me look into it. Might not be worth the overhead just to use Mixture, but I will see if that's true.

fonnesbeck · 2017-05-29T14:29:36Z

Actually, its pretty easy. Hold off on merging this for the moment; I will convert all the zero-inflated distributions over and move them to distributions/mixture.

twiecki · 2017-05-29T14:37:45Z

Is it? I think we tried it but it wasn't easy to define a 0-distribution. CC @AustinRochford

fonnesbeck · 2017-05-29T14:39:11Z

Just used a Constant.dist(0).

Here is the zero-inflated Poisson in action.

twiecki · 2017-05-29T14:39:13Z

#1459

fonnesbeck · 2017-05-29T14:48:50Z

And here is zero-inflated binomial. Seems to work. Let me know if I am missing something (I very well could be).

fonnesbeck · 2017-05-29T14:51:50Z

Please test.

junpenglao · 2017-05-29T15:25:34Z

pymc3/distributions/mixture.py

+    """
+
+    def __init__(self, psi, theta, *args, **kwargs):
+        self.theta = theta = tt.as_tensor_variable(theta)


you should also cast psi into tensor (same below), although might not necessary after #2233 is merge

That happens in the superclass init (Line 48)

Fixing typo

…classes

fonnesbeck · 2017-05-29T23:07:09Z

Since these are always 2-component mixtures, I've changed the psi to import a scalar rather than a vector (e.g. a Beta rather than a Dirichlet). This is consistent with the API of the existing zero-inflateds.

…ion of psi in docstring

fonnesbeck · 2017-05-30T03:34:24Z

Seems to have a problem generating random draws with a multidimensional size. Not sure why.

AustinRochford · 2017-05-30T12:03:25Z

Looks good overall. I had similar problems matching scipy and with random draws when I tried this before. Will give some thought as to possible problems.

* Change sample() to use live_plot_kwargs instead of **kwargs. * Pass kwargs of sample to _sample. * Typo wargs -> kwargs. * Re-add kwargs to _sample().

* [WIP] clean up GLM doc * pm.glm.glm -> pm.glm.GLM.from_formula * pm.glm.plot_posterior_predictive -> pm.plot_posterior_predictive_glm * remove find_MAP and used default init * More clean up * more clean up * GLM doc clean up last notebook. change to a slidely different example.

AustinRochford · 2017-05-31T00:04:24Z

Passing broadcast_conditions=False to bound in Mixture's logp method fixes the broken ZIP and ZINB tests for me, but the ZIB test is still failing.

This was the reason for broadcast_conditions, so I'll add that in a separate PR and keep investigating the ZIB failure.

AustinRochford · 2017-05-31T00:13:42Z

pymc3/distributions/mixture.py

+    Parameters
+    ----------
+    psi : float
+        Expected proportion of Poisson variates (0 < psi < 1)


Should say Binomial

This reverts commit 482e751.

…cast-conditions Don't broadcast conditions in Mixture logp

AustinRochford · 2017-05-31T02:00:09Z

@fonnesbeck the broadcast_conditions change is merged to master, rebasing should get more (but not all) tests to pass.

…nent mixture

…classes

…ion of psi in docstring

…nto zero_inflated_binomial

fonnesbeck · 2017-05-31T03:12:23Z

Well, I rebased instead of merged, so this branch is irrevocably messed up. Closing in favor of #2246

fonnesbeck and others added 16 commits May 27, 2017 14:46

Added zero-inflated binomial distribution

b407bdb

Fixed syntax error

524b88e

Added ZeroInflatedBinomial to global namespace

17b27b0

Fixed attribute error in call to scipy.stats

4fd07b8

Use setup.py to create test environment

da5401f

Mock is also used

cfd29aa

Imported but untested requirements

d35f0e5

Use conda numpy/scipy/mkl-service

0f7f6b5

Merge pull request #2229 from ColCarroll/update_create_testenv

b505f47

Update create testenv

Vi summary (#2230)

c3afa00

* add draft * add draft * move and rename * better docstring * added tests

Return identity matrix if no scaling provided

52761c2

Simplify calculation of identity scaling matrix

c4659b4

Merge pull request #2232 from a-rodin/scaling-default-to-identity

eb6042f

WIP: replace Hessian scaling guessing by an identity matrix

scaling for NUTS using 'advi_map'

4a8f64f

remaining fix in #2110

Revert "Merge pull request #2232 from a-rodin/scaling-default-to-iden…

1aea645

…tity" This reverts commit eb6042f, reversing changes made to c3afa00.

junpenglao reviewed May 29, 2017

View reviewed changes

Junpeng Lao and others added 2 commits May 29, 2017 15:54

Merge pull request #2234 from junpenglao/master

9b0f4fc

scaling for NUTS using 'advi_map'

Imported ZeroInflatedBinomial in test_distributions

999be1e

Converted zero-inflated distributions to Mixture subclasses

e478fe3

fonnesbeck changed the title ~~Zero-inflated Binomial~~ Convert zero-inflated distributions to Mixture subclasses May 29, 2017

junpenglao reviewed May 29, 2017

View reviewed changes

Junpeng Lao and others added 2 commits May 29, 2017 22:36

Merge pull request #2237 from karink520/mixture_typo

69e223b

Fixing typo

Fixed dimension-matching issue with mixture weights in zero-inflated …

a750e3d

…classes

Reversed order of psi, 1-psi to make them consistent with the definit…

4515977

…ion of psi in docstring

twiecki and others added 2 commits May 30, 2017 17:06

Change sample() to use live_plot_kwargs instead of **kwargs. (#2235)

3d83c6d

* Change sample() to use live_plot_kwargs instead of **kwargs. * Pass kwargs of sample to _sample. * Typo wargs -> kwargs. * Re-add kwargs to _sample().

Don't broadcast conditions in Mixture logp

9cb6d53

AustinRochford mentioned this pull request May 31, 2017

Don't broadcast conditions in Mixture logp #2243

Merged

AustinRochford reviewed May 31, 2017

View reviewed changes

AustinRochford added 3 commits May 30, 2017 21:18

statsmodels is required for some examples

482e751

Revert "statsmodels is required for some examples"

9047dfa

This reverts commit 482e751.

Merge pull request #2243 from AustinRochford/mixture-bound-dont-broad…

541afd8

…cast-conditions Don't broadcast conditions in Mixture logp

fonnesbeck added 10 commits May 30, 2017 21:55

Added zero-inflated binomial distribution

504fd60

Fixed syntax error

8a9936b

Added ZeroInflatedBinomial to global namespace

9f7e8ee

Fixed attribute error in call to scipy.stats

16576f9

Imported ZeroInflatedBinomial in test_distributions

b7cfd1f

Converted zero-inflated distributions to Mixture subclasses

b4494c6

Changed docstring of zero-inflated distributions to specify two-compo…

e847161

…nent mixture

Fixed dimension-matching issue with mixture weights in zero-inflated …

4e8df1c

…classes

Reversed order of psi, 1-psi to make them consistent with the definit…

1e05632

…ion of psi in docstring

Merge branch 'zero_inflated_binomial' of github.com:pymc-devs/pymc3 i…

0217259

…nto zero_inflated_binomial

fonnesbeck mentioned this pull request May 31, 2017

WIP: Convert zero-inflated distributions to Mixture subclasses #2246

Closed

fonnesbeck closed this May 31, 2017

junpenglao deleted the zero_inflated_binomial branch May 31, 2017 05:21

Uh oh!

Convert zero-inflated distributions to Mixture subclasses #2231

Convert zero-inflated distributions to Mixture subclasses #2231

Uh oh!

Conversation

fonnesbeck commented May 27, 2017

Uh oh!

junpenglao May 29, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

twiecki commented May 29, 2017

Uh oh!

fonnesbeck commented May 29, 2017

Uh oh!

fonnesbeck commented May 29, 2017

Uh oh!

twiecki commented May 29, 2017

Uh oh!

fonnesbeck commented May 29, 2017

Uh oh!

twiecki commented May 29, 2017

Uh oh!

fonnesbeck commented May 29, 2017

Uh oh!

fonnesbeck commented May 29, 2017

Uh oh!

junpenglao May 29, 2017

Choose a reason for hiding this comment

Uh oh!

fonnesbeck May 29, 2017

Choose a reason for hiding this comment

Uh oh!

fonnesbeck commented May 29, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fonnesbeck commented May 30, 2017

Uh oh!

AustinRochford commented May 30, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AustinRochford commented May 31, 2017

Uh oh!

AustinRochford May 31, 2017

Choose a reason for hiding this comment

Uh oh!

AustinRochford commented May 31, 2017

Uh oh!

fonnesbeck commented May 31, 2017

Uh oh!

Uh oh!

junpenglao May 29, 2017 •

edited

Loading

fonnesbeck commented May 29, 2017 •

edited

Loading

AustinRochford commented May 30, 2017 •

edited

Loading