Added base class for variational methods #1600

ferrine · 2016-12-13T16:48:50Z

Because of upcomming Normalizing flows and Stein variational gradient descent I propose a unified interface for variational methods

…tting (e.g. nanguardmode) for Theano functions

* Started to write Base class for pymc3.models * mode `add_var` to public api * Added some docstrings * Added some docstrings * added getitem and fixed a typo * added assertion check * added resolve var method * decided not to add resolve method * Added linear component * Docs fix * patsy's intercept is inited properly now * refactored code * updated docs * added possibility to init coefficients with random variables * added glm * refactored api, fixed formula init * refactored linear model, extended acceptable types * moved useful matrix and labels creation to utils file * code style * removed redundant evaluation of shape * refactored resolver for constructing matrix and labels * changed error message * changed signature of init * simplified utils any_to_tensor_and_labels code * tests for `any_to_tensor_and_labels` * added docstring for `any_to_tensor_and_labels` util * forgot to document return type in `any_to_tensor_and_labels` * refactored code for dict * dict tests fix(do not check labels there) * added access to random vars of model * added a shortcut for all variables so there is a unified way to get them * added default priors for linear model * update docs for linear * refactored UserModel api, made it more similar to pm.Model class * Lots of refactoring, tests for base class, more plain api design * deleted unused module variable * fixed some typos in docstring * Refactored pm.Model class, now it is ready for inheritance * Added documentation for Model class * Small typo in docstring * nested contains for treedict (needed for add_random_variable) * More accurate duplicate implementation of treedict/treelist * refactored treedict/treelist * changed `__imul__` of treelist * added `root` property and `isroot` indicator for base model * protect `parent` and `model` attributes from violation * travis' python2 did not fail on bad syntax(maybe it's too new), fixed * decided not to use functools wrapper Unfortunately functools wrapper fails when decorating built-in methods so I need to fix that improper behaviour. Some bad but needed tricks were implemented * Added models package to setup script * Refactor utils * Fix some typos in pm.model

…tting (e.g. nanguardmode) for Theano functions

* Started to write Base class for pymc3.models * mode `add_var` to public api * Added some docstrings * Added some docstrings * added getitem and fixed a typo * added assertion check * added resolve var method * decided not to add resolve method * Added linear component * Docs fix * patsy's intercept is inited properly now * refactored code * updated docs * added possibility to init coefficients with random variables * added glm * refactored api, fixed formula init * refactored linear model, extended acceptable types * moved useful matrix and labels creation to utils file * code style * removed redundant evaluation of shape * refactored resolver for constructing matrix and labels * changed error message * changed signature of init * simplified utils any_to_tensor_and_labels code * tests for `any_to_tensor_and_labels` * added docstring for `any_to_tensor_and_labels` util * forgot to document return type in `any_to_tensor_and_labels` * refactored code for dict * dict tests fix(do not check labels there) * added access to random vars of model * added a shortcut for all variables so there is a unified way to get them * added default priors for linear model * update docs for linear * refactored UserModel api, made it more similar to pm.Model class * Lots of refactoring, tests for base class, more plain api design * deleted unused module variable * fixed some typos in docstring * Refactored pm.Model class, now it is ready for inheritance * Added documentation for Model class * Small typo in docstring * nested contains for treedict (needed for add_random_variable) * More accurate duplicate implementation of treedict/treelist * refactored treedict/treelist * changed `__imul__` of treelist * added `root` property and `isroot` indicator for base model * protect `parent` and `model` attributes from violation * travis' python2 did not fail on bad syntax(maybe it's too new), fixed * decided not to use functools wrapper Unfortunately functools wrapper fails when decorating built-in methods so I need to fix that improper behaviour. Some bad but needed tricks were implemented * Added models package to setup script * Refactor utils * Fix some typos in pm.model

twiecki · 2016-12-13T17:15:19Z

Seems like you need to rebase.

# Conflicts: # pymc3/model.py # pymc3/step_methods/nuts.py

ferrine · 2016-12-15T18:45:01Z

After some tests it will be ready for merge, so I need a review

* refactored GARCH and added Mv(Gaussian/StudentT)RandomWalk * refactored garch logp * added docs * fix typo * even more typos * better description for tau

springcoil · 2016-12-15T23:04:05Z

pymc3/math.py

+
+def flatten(tensors, outdim=1):
+    if not isinstance(tensors, list):
+        return _flatten(tensors, outdim)


Hmm using a private method that's the same name as the other one. I'm not so sure about this.

Actually you've done it quite a lot - so this is just a style thing - I don't see a huge problem with it.

I was thinking about it too. It should be better named as list_flatten or something like this

springcoil · 2016-12-15T23:06:04Z

pymc3/variational/replacements.py

+        samples = tt.as_tensor(samples)
+        pi = tt.as_tensor(pi)
+        _elbo_ = self.log_p_D + pi * (self.log_p_W - self.log_q_W)
+        elbos, updates = theano.scan(fn=lambda: _elbo_,


I like this

Good use of the theano.scan functionality, and hidden well in the API.

springcoil · 2016-12-15T23:07:36Z

Code looks clean - a few stylistic concerns but nothing major. @ferrine do you want another review before this merges?

ferrine · 2016-12-15T23:11:57Z

Thanks, @springcoil. Yes, I want some more reviews as I want Normalizing Flows and SVGD to be implemented through this unified api. CC @twiecki @taku-y

springcoil · 2017-01-04T22:27:57Z

Cool. I'd say split this into two PR's it just makes review so much easier :)

…tting (e.g. nanguardmode) for Theano functions

… mode setting (e.g. nanguardmode) for Theano functions" This reverts commit 6bdb53a.

# Conflicts: # pymc3/distributions/dist_math.py # pymc3/distributions/distribution.py # pymc3/model.py # pymc3/step_methods/hmc/nuts.py # pymc3/step_methods/hmc/trajectory.py # pymc3/tests/test_step.py # pymc3/theanof.py # pymc3/variational/advi.py # pymc3/variational/advi_minibatch.py # setup.py

ferrine · 2017-01-17T22:58:30Z

That was pretty bad merge... So, I've tested scaling and we can ignore scaling densities in the future. This PR probably needs hard refactoring again.

ferrine · 2017-01-17T23:39:19Z

Hmm, I think we are able to add recompution of models logp so that it will be evaluated on a number of samples. This way we can avoid costly scan op when sampling logp. It will extend Model's interface btw. I'll add it to PR

ferrine · 2017-01-17T23:59:44Z

I'm planning to refactor most of the code as I'm not satisfied with approach I use now. I'll do advi in OPVI framework soon. Implementing LG opvi will be in a couple of lines as well as advi. Here is a draft for supposed opvi interface

taku-y · 2017-01-18T01:19:11Z

Sounds great. Hope your refactoring goes well. Btw, I'm curious about removing scan. Is it related to the shape of RV (#1125)?

ferrine · 2017-01-18T08:17:03Z

@taku-y, no, it's not related as I want to do elemwise computation on custom tensor. I'm afraid how it will look like on Multi Observed RV

taku-y · 2017-01-18T10:18:15Z

@ferrine Thanks for your reply. I shuold wait your code. I look forward to your new framework!

ferrine · 2017-01-19T21:59:49Z

I've tried to rebase the branch on master but it was full of pain. I would rather create a new branch and OPVI PR. I have lot's of things to refactor so there will be a mess soon otherwise

ferrine · 2017-01-22T12:37:16Z

Closing in favour of #1694

fonnesbeck and others added 20 commits November 28, 2016 05:36

Added mode argument to several step methods and advi to allow mode se…

2d6fee8

…tting (e.g. nanguardmode) for Theano functions

Fixed namespace bugs in mode attribute

4e5b9c2

Reverted function in delta_logp to not accept mode argument

0ebaacd

added new elbo implementation

9ab04da

Added mode argument to several step methods and advi to allow mode se…

9811220

…tting (e.g. nanguardmode) for Theano functions

Fixed namespace bugs in mode attribute

fbd1d5b

Reverted function in delta_logp to not accept mode argument

208aa79

Refactor Hamiltonian methods into single class

ea82ebd

Reformat docs

140a80c

added replacements class and mean field approximation

168b113

moved local to local constructor

c1211a6

property for deterministic replacements

9690562

refactored replacements to make them more unitary

34da7c8

shape problem when sampling

07a248a

tests passed

889b50e

deleted unused modules

0d486fb

added replacement names for global/local dict

125f6ad

ferrine added 3 commits December 13, 2016 20:34

Merge branch '3.1' into refactor_advi

1af91c0

# Conflicts: # pymc3/model.py # pymc3/step_methods/nuts.py

refactored replacements

69f07a1

refactored replacements

9614bf9

ferrine added 2 commits December 16, 2016 01:40

refactored GARCH and added Mv(Gaussian/StudentT)RandomWalk (#1603)

32a2eb7

* refactored GARCH and added Mv(Gaussian/StudentT)RandomWalk * refactored garch logp * added docs * fix typo * even more typos * better description for tau

Merge branch '3.1' into refactor_advi

5e68b95

springcoil reviewed Dec 15, 2016

View reviewed changes

twiecki force-pushed the 3.1 branch from 7053e43 to 4cef603 Compare January 9, 2017 16:07

fonnesbeck and others added 17 commits January 14, 2017 14:55

Added mode argument to several step methods and advi to allow mode se…

8d1f088

…tting (e.g. nanguardmode) for Theano functions

Created Generator Op with simple Test

37843af

added ndim test

3dc6f1b

updated test

a16512e

updated test, added test value check

7127c23

added test for replacing generator with shared variable

23b14ff

added shortcut for generator op

96cd5bb

refactored test

633e4e9

added population kwarg (no tests yet)

0629adc

added population kwarg for free var(autoencoder case)

06099a2

Revert "Added mode argument to several step methods and advi to allow…

75a4849

… mode setting (e.g. nanguardmode) for Theano functions" This reverts commit 6bdb53a.

add docstring to generator Op

ff325d8

rename population -> total_size

79ac934

update docstrings in model

57dbe47

fix typo in as_tensor function

f8bce58

add simple test for density scaling via total_size

8d91fee

raise an error when density scaling is done on scalar

1a9fa3d

ferrine closed this Jan 22, 2017

ferrine mentioned this pull request Jan 22, 2017

WIP: Implement opvi #1694

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Added base class for variational methods #1600

Added base class for variational methods #1600

ferrine commented Dec 13, 2016 •

edited by fonnesbeck

Loading

twiecki commented Dec 13, 2016

ferrine commented Dec 15, 2016

springcoil Dec 15, 2016

springcoil Dec 15, 2016

ferrine Dec 15, 2016

springcoil Dec 15, 2016

springcoil Dec 15, 2016

springcoil commented Dec 15, 2016

ferrine commented Dec 15, 2016

springcoil commented Jan 4, 2017

ferrine commented Jan 17, 2017

ferrine commented Jan 17, 2017 •

edited

Loading

ferrine commented Jan 17, 2017

taku-y commented Jan 18, 2017

ferrine commented Jan 18, 2017

taku-y commented Jan 18, 2017

ferrine commented Jan 19, 2017

ferrine commented Jan 22, 2017

Added base class for variational methods #1600

Added base class for variational methods #1600

Conversation

ferrine commented Dec 13, 2016 • edited by fonnesbeck Loading

twiecki commented Dec 13, 2016

ferrine commented Dec 15, 2016

springcoil Dec 15, 2016

Choose a reason for hiding this comment

springcoil Dec 15, 2016

Choose a reason for hiding this comment

ferrine Dec 15, 2016

Choose a reason for hiding this comment

springcoil Dec 15, 2016

Choose a reason for hiding this comment

springcoil Dec 15, 2016

Choose a reason for hiding this comment

springcoil commented Dec 15, 2016

ferrine commented Dec 15, 2016

springcoil commented Jan 4, 2017

ferrine commented Jan 17, 2017

ferrine commented Jan 17, 2017 • edited Loading

ferrine commented Jan 17, 2017

taku-y commented Jan 18, 2017

ferrine commented Jan 18, 2017

taku-y commented Jan 18, 2017

ferrine commented Jan 19, 2017

ferrine commented Jan 22, 2017

ferrine commented Dec 13, 2016 •

edited by fonnesbeck

Loading

ferrine commented Jan 17, 2017 •

edited

Loading