Maint: Hypervolume (pyhv.py) module rewrite #484

Corwinpro · 2020-01-29T13:39:06Z

Types of changes

Docs change / refactoring / dependency upgrade
Bug fix (non-breaking change which fixes an issue)
New feature (non-breaking change which adds functionality)
Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

This PR is related to the #366.

This PR rewrites the code that generates the hypervolume indicator, used for multiobjective optimization.

Any suggestions / comments / requests are very welcome.

The purpose is two-fold:

Rewrite / refactor the existing code to improve the general readability and extendability, spot possible bugs, and add documentation
Get rid of the LGPL licence in favour of the standard MIT licence, used by nevergrad.

How Has This Been Tested (if it applies)

We added unit tests for the data structure that are used by the HypervolumeIndicator class.

We test the main functionality of the new hypervolume algorithm against the old one. More integration tests should be included before this PR can be accepted.

Checklist

The documentation is up-to-date with the changes I made.
I have read the CONTRIBUTING document and completed the CLA (see CONTRIBUTING).
All tests passed, and additional code has been covered with new tests.

teytaud · 2020-01-29T13:43:35Z

Waow thanks!
I try to read this promptly - what is sure is that this was much needed!

Corwinpro · 2020-01-29T13:48:28Z

Thank you @teytaud

I am fixing the mypy errors now, although it can take some time as I have never used this type checking system before.

teytaud · 2020-01-30T07:43:37Z

Congratulations, all the real tests work! The only issue is the Type information: mypy detects quite a bit of type info missing. Can you do that ?

Corwinpro · 2020-01-30T09:47:23Z

Thant you @teytaud . Yes, I will be adding the Type info shortly: mypy is a new system for me, so I apologize in advance for the possible delays.

jrapin

I've not dived into the actual implementation but here are a few initial feedbacks.
Replacing pyhv by this in the rest of the code will help understand it better and check for bugs.
In the meantime, do you want me to make a PR on your PR to help you add typing?

nevergrad/functions/multiobjective/tests/test_structures.py

nevergrad/functions/multiobjective/hypervolume.py

Corwinpro · 2020-01-30T13:15:43Z

Thank you @jrapin ! I am fine with using the same branch if you would like to - I can focus on the testing task if you prefer.

…grad into enthought-maint/pyhv-rewrite

jrapin · 2020-01-31T16:09:49Z

Thank you @jrapin ! I am fine with using the same branch if you would like to - I can focus on the testing task if you prefer.

I tend to prefer PRs so as to highlight changes, and avoid collisions. However I don't seem to be able to PR on your fork, I don't understand why :s so you can find typing that works on branch enthought-maint/pyhv-rewrite-typing
I went to the simplest solutions to make it work, sometimes ignoring the typing, and other times using Any when I did not understand it.
Note that there are a few unbound variables (see TODOs)

Edit: doing this it seems I have broken the code though. The "simplest solution" was to assume that next and prev never contained Nones, but it seems it needs to. The issue then is that most of the time it is assumed that they are not None. I'm a bit concerned by this pattern, whenever possible we should work with non "Optional" data, that simplifies the typing, and also avoids weird bugs

Corwinpro · 2020-01-31T16:42:43Z

Thank you @jrapin for this PR, it is great. Sorry for the delay, I might be able to finish the testing by Monday night.

Corwinpro · 2020-01-31T16:53:21Z

I guess node.next[i] can be None - this is the case when we have a default node, with no connections.

Regarding the TODOs, I am updating this comment with what I think can be useful:

The HypervolumeIndicator.compute is the main method of the algorigthm:

    def compute(self, points: tp.Any) -> float:  # TODO not too sure what that is (replace Any)

It is used by MultiobjectiveFunction.compute_aggregate_loss (line 52 in functions/multiobjective/core.py. I would assume that a list of "vectors" would be passed as the argument.

…n-1)

Corwinpro · 2020-02-03T12:44:27Z

I believe the mypy issues should be resolved now.

HypervolumeIndicator tests should be on the way, working on this now.

…ervolumeIndicator

Corwinpro · 2020-02-04T16:42:42Z

@jrapin Could you please have a look at this now? Most of the tests are ready, I guess it could be a good time for review and corrections, if you don't mind.

Thank you!

jrapin

Sorry for the delay.
Here are a few comments.
I checked mostly the code, but as long as the tests are working (and they are) and with a few minor updates (mostly parameters that I feel are not needed) I am OK.
I'll probably iterate on this in (much) later PRs, I feel there is some underlying structure that may be made more explicit, but dunno what exactly for now anyway.

jrapin · 2020-02-05T15:49:11Z

nevergrad/functions/multiobjective/hypervolume.py

+        hypervolume += h * last_node.coordinate[dimension]
+        return hypervolume
+
+    def recursive_hypervolume(self, dimension: int, bounds: np.ndarray) -> float:


Is there any reason for this method to take dimension and bounds as inputs? Those seem already provided by the instance

Thank you, I totally agree that the bounds are passed in for now reason. Should be fixed in the next commit.

The dimension though is needed, as we recursively iterate from the default dimension of the problem to the lower dimensions; it doesn't always equal the dimension of the space of the problem.

jrapin · 2020-02-05T15:50:57Z

nevergrad/functions/multiobjective/hypervolume.py

+        hypervolume = self.recursive_hypervolume(self.dimension - 1, self.reference_bounds)
+        return hypervolume
+
+    def plane_hypervolume(self, dimension: int) -> float:


dimension seems provided by the instance and necessarily 1, so you could remove this and if need be,
check self.dimension == 2?

If I am not mistaken, the self.dimension should not necessarily be 2 here: as we iterate from the higher to lower dimensional subspaces, at some point we will reach the space of dim == 2. Then, this method is called.

I removed the argument dimension from the method - it was ambiguous.

Thank you.

jrapin · 2020-02-05T15:52:27Z

nevergrad/functions/multiobjective/hypervolume.py

+        hypervolume -= last_node.area[dimension] * last_node.coordinate[dimension]
+        return hypervolume
+
+    def skip_dominated_points(


Same here, I would say:

either the methods does not need the argument,

or it should be a function, not a method

But maybe I'm not understanding it properly

I have the same reasoning as before - the dimension will change during the recursive iteration from the higher to lower dimensional space. I hope I understand this algorithm correctly; I will double check that.

nevergrad/functions/multiobjective/hypervolume.py

jrapin · 2020-02-05T15:57:53Z

nevergrad/functions/multiobjective/hypervolume.py

+import numpy as np
+
+
+class VectorNode:


Can you add some docstring here, just to give a brief overview of what this does, and how it is useful ? + describe the parameters.
Btw, shouldnt it be coordinates and not coordinate? I'm not a native English speaker though so I may be wrong

Correct. Fixed that.

nevergrad/functions/multiobjective/hypervolume.py

jrapin · 2020-02-05T16:05:25Z

@teytaud anything to add?
@Corwinpro You can probably remove the pyhv file and the related license

…; using the instances .reference_bounds array instead

Corwinpro · 2020-02-07T15:11:38Z

I removed the old code and the license file. Would you like me to add anything else?

Thank you!

jrapin

I forgot all the licensing stuff, see comments below + can you remove the sentence " LGPL code is however also included in the multiobjective subpackage." in the README?
@teytaud anything more to say?
I didn't have time to dive into the algorithm but since tests are passing it's perfect for me. Thanks for the help!

jrapin · 2020-02-10T09:08:39Z

nevergrad/functions/multiobjective/hypervolume.py

@@ -0,0 +1,303 @@
+import typing as tp


Can you add the same header as on all other files? This is required to mention that it is released under the MIT license

jrapin · 2020-02-10T09:22:16Z

nevergrad/functions/multiobjective/test_structures.py

@@ -0,0 +1,304 @@
+import numpy as np


Please add the header here too

Corwinpro · 2020-02-11T16:51:03Z

Thank you very much for the review and the comments. I am ready to merge this if you are happy with all the changes.

Thank you!

jrapin · 2020-02-12T09:54:17Z

Sorry for the delay, let's go then ;)

Petr Kungurtsev added 3 commits January 28, 2020 14:58

DEV: First attempt to implement the Hypervolume Indicator from the paper

1ef9ffa

Merge branch 'master' into maint/pyhv-rewrite

189f5e7

MAINT: string representation of the VectorLinkedList

fa9d3aa

facebook-github-bot added the CLA Signed Do not delete this pull request or issue due to inactivity. label Jan 29, 2020

jrapin reviewed Jan 30, 2020

View reviewed changes

Petr Kungurtsev and others added 5 commits January 30, 2020 13:17

MAINT: Node.pop method moved to the class

b4e1daf

TST: ported tests to pytest

a00d5d9

TST: moved test file to the root folder

8944689

Merge branch 'maint/pyhv-rewrite' of git://github.com/enthought/never…

81399b7

…grad into enthought-maint/pyhv-rewrite

Update typing

cffd0a1

Petr Kungurtsev added 7 commits January 31, 2020 17:27

FIX: default node prev and next

c22e01b

FIX: default node prev and next

b92457a

DEV: added length of multilist in i-th index

13a87cf

MAINT: undefined reference fixed

499e414

DEV: front_size is equivalent to self.multilist.chain_length(dimensio…

5587248

…n-1)

MAINT: fixed refernces that can possibly be undefined

1625177

MAINT: node's prev and next fix, including mypy typing

fdade40

Petr Kungurtsev added 4 commits February 3, 2020 12:54

TST: Test fixes

74429ed

TST: integration test moved to the test file

1db801a

TST: Tests for Node and VectorList structures

39afe74

Merge branch 'master' into maint/pyhv-rewrite

58be6ff

Petr Kungurtsev added 8 commits February 3, 2020 13:58

MAINT: removed refernce to old _HyperVolume in core

f1acdb6

MAINT: moved the VectorLinkedList methods to that class, from the Hyp…

f35a907

…ervolumeIndicator

STY: minor typos fix

33d6093

TST: Test updates for VectorLinkedList

eec464f

MAINT: mypy updates

a617603

MAINT: changing type List[float] to ndarray in bounds

aca1ef7

MAINT: mypy changes to type of compute() arguments

3edfca6

FIX: np.array type (should've been np.ndarray)

fa72612

Merge branch 'master' into maint/pyhv-rewrite

4c955fe

jrapin reviewed Feb 5, 2020

View reviewed changes

Petr Kungurtsev added 6 commits February 5, 2020 19:10

Merge branch 'master' into maint/pyhv-rewrite

7d29aaa

MAINT: remove bounds from the recursive_hypervolume list of arguments…

e3f9fd4

…; using the instances .reference_bounds array instead

MAINT: removed dimension argument from the plane hypervolume

8d42854

MAINT: chain_length suggestion

8c1c2f6

MAINT: docstring added and attribute coordinate renamed

73ed35c

MAINT: removed old code

9487e71

jrapin approved these changes Feb 10, 2020

View reviewed changes

Petr Kungurtsev added 3 commits February 10, 2020 09:47

Merge branch 'master' into maint/pyhv-rewrite

563f20e

DOC: Removed LGPL comment in README

3b93968

LICENCE: Added MIT Licence headers to new files

68f4c82

jrapin merged commit caf2059 into facebookresearch:master Feb 12, 2020

Corwinpro deleted the maint/pyhv-rewrite branch February 12, 2020 09:58

Corwinpro mentioned this pull request Feb 19, 2020

Missing EU project reference in the merged PR #529

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Maint: Hypervolume (pyhv.py) module rewrite #484

Maint: Hypervolume (pyhv.py) module rewrite #484

Corwinpro commented Jan 29, 2020 •

edited

Loading

teytaud commented Jan 29, 2020

Corwinpro commented Jan 29, 2020

teytaud commented Jan 30, 2020

Corwinpro commented Jan 30, 2020

jrapin left a comment

Corwinpro commented Jan 30, 2020

jrapin commented Jan 31, 2020 •

edited

Loading

Corwinpro commented Jan 31, 2020 •

edited

Loading

Corwinpro commented Jan 31, 2020 •

edited

Loading

Corwinpro commented Feb 3, 2020 •

edited

Loading

Corwinpro commented Feb 4, 2020

jrapin left a comment

jrapin Feb 5, 2020

Corwinpro Feb 5, 2020

jrapin Feb 5, 2020

Corwinpro Feb 5, 2020

jrapin Feb 5, 2020

Corwinpro Feb 5, 2020

jrapin Feb 5, 2020

Corwinpro Feb 5, 2020 •

edited

Loading

jrapin commented Feb 5, 2020

Corwinpro commented Feb 7, 2020 •

edited

Loading

jrapin left a comment

jrapin Feb 10, 2020

jrapin Feb 10, 2020

Corwinpro commented Feb 11, 2020

jrapin commented Feb 12, 2020

		import numpy as np


		class VectorNode:

Maint: Hypervolume (pyhv.py) module rewrite #484

Maint: Hypervolume (pyhv.py) module rewrite #484

Conversation

Corwinpro commented Jan 29, 2020 • edited Loading

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

teytaud commented Jan 29, 2020

Corwinpro commented Jan 29, 2020

teytaud commented Jan 30, 2020

Corwinpro commented Jan 30, 2020

jrapin left a comment

Choose a reason for hiding this comment

Corwinpro commented Jan 30, 2020

jrapin commented Jan 31, 2020 • edited Loading

Corwinpro commented Jan 31, 2020 • edited Loading

Corwinpro commented Jan 31, 2020 • edited Loading

Corwinpro commented Feb 3, 2020 • edited Loading

Corwinpro commented Feb 4, 2020

jrapin left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Corwinpro Feb 5, 2020 • edited Loading

Choose a reason for hiding this comment

jrapin commented Feb 5, 2020

Corwinpro commented Feb 7, 2020 • edited Loading

jrapin left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Corwinpro commented Feb 11, 2020

jrapin commented Feb 12, 2020

Corwinpro commented Jan 29, 2020 •

edited

Loading

jrapin commented Jan 31, 2020 •

edited

Loading

Corwinpro commented Jan 31, 2020 •

edited

Loading

Corwinpro commented Jan 31, 2020 •

edited

Loading

Corwinpro commented Feb 3, 2020 •

edited

Loading

Corwinpro Feb 5, 2020 •

edited

Loading

Corwinpro commented Feb 7, 2020 •

edited

Loading