Add cost function decorator #226

whzup · 2018-08-15T17:52:31Z

Description

This decorator allows the creation of much simpler cost functions. Instead of writing a cost function that returns a shape of (n_particles, 0) it enables the usage of shorter and simpler cost functions that directly return the cost.

Related Issue

Resolves: #228

Motivation and Context

At the moment, the way one has to write the cost function is very counterintuitive as it is necessary to write a function that returns an array of shape (n_particles, ). This decorator "vectorizes" a cost function which returns just the cost.

How Has This Been Tested?

Wrote a new test file.

Types of changes

Bug fix (a non-breaking change which fixes an issue)
New feature (a non-breaking change which adds functionality)
Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

My code follows the code style of this project.
My change requires a change to the documentation.
I have updated the documentation accordingly.
I have read the CONTRIBUTING document.
I have added tests to cover my changes.
All new and existing tests pass.

ljvmiranda921 · 2018-08-17T00:18:27Z

pyswarms/utils/decorators/decorators.py

+    :code:`int`.
+
+    Attributes
+    ==========


I think this should be a second-level header, and Parameters:

def cost(cost_func): """ Parameters ---------- """

Same with Returns 👍

Yes, thanks for the catch 😄

ljvmiranda921 · 2018-08-17T00:19:30Z

I'll update this branch first (not sure if you can click the button over there) so that #229 will also reflect here:

ljvmiranda921 · 2018-08-17T07:20:40Z

tests/utils/decorators/test_decorators.py

+
+@pytest.mark.parametrize(
+    "objective_func",
+    [np.sum(), np.prod()]


Just pass the function reference: np.sum and np.prod

Can you also try np.dot? Say you're given a particle with d dimensions, and you dot-product (sumproduct) it to an objective function (randomly-generated array with d dimensions), you only get 1 output. With the @cost decorator, you can run this on a set of n particles with d dimensions too 👍 This workflow is similar to our neural network example, so we might as well try 👍

I'd have to define a new function for this since np.dot takes two arguments. Shall I put it in the conftest.py file?

ljvmiranda921 · 2018-08-17T07:21:39Z

pyswarms/utils/decorators/decorators.py

+    Parameters
+    ----------
+
+    cost_func : function


This should be callable instead of function

ljvmiranda921 · 2018-08-17T07:22:32Z

pyswarms/utils/decorators/decorators.py

+    ----------
+
+    cost_func : function
+        A function that can be used as cost function in the optimization (must return a :code:`float` or an :code:`int`)


This line's a bit long (?) I don't think black formats docstrings, can flake8 catch this line? What does flake8 decorators.py say?

ljvmiranda921 · 2018-08-17T07:27:09Z

tests/utils/decorators/test_decorators.py

+    w_o_decorator = cost_func_w_o_decorator(particles)
+    w_decorator = cost_func_w_decorator(particles)
+
+    assert isinstance(w_decorator, Number)


Hmmm, I'm not sure why we need to test if the output is a Number. Maybe we should test the shape of the resulting array? Ensure that it's (n_particles, )?

ljvmiranda921 · 2018-08-17T07:27:33Z

pyswarms/utils/decorators/decorators.py

+    -------
+
+    cost_dec : function
+        The vectorized output for all particles as defined by :code:`cost_func`


This should be:

""" Returns ---------- callable Put description here """

ljvmiranda921 · 2018-08-17T07:39:14Z

pyswarms/utils/decorators/decorators.py

+    needs to resemble a strategy that can be applied to every particle in the swarm
+    to assign a cost. This means that the decorated cost function must use an array
+    of shape :code:`(dimensions, )` as an argument and return a :code:`float` or an
+    :code:`int`.


I think this can be phrased as:

If you have defined a cost function that only acts upon a single particle, then you can use this decorator to have a vectorized computation for n particles. ...

Then give an example code-block maybe? I think good keywords here are: wrapper, transforms, etc. Strategy seems a bit odd (I think you got this from hypothesis haha but I think it's more of their local lingo).

ljvmiranda921 · 2018-08-17T07:49:20Z

tests/utils/decorators/test_decorators.py

+    "objective_func",
+    [np.sum(), np.prod()]
+)
+def test_cost_decorator(objective_func, particles):


Where do you get the particles? I suggest you can generate them via np.random.uniform...

Do you know a way we can output the seed number with the tests without using pytest -s? In this way, we don't have fixed particle arrays and if it fails we can use the printed seed to debug it.

ljvmiranda921 · 2018-08-17T07:50:21Z

tests/utils/decorators/test_decorators.py

+    [np.sum(), np.prod()]
+)
+def test_cost_decorator(objective_func, particles):
+    def cost_func_w_o_decorator(x):


It's ok to spell it out 😆 cost_func_without_decorator(x) and cost_fun_with_decorator

ljvmiranda921 · 2018-08-17T08:00:20Z

Aside from my comments, I think this is great, @whzup ! This is an elegant and useful improvement to our current codebase! Very excited to try this out!

whzup · 2018-08-17T15:44:14Z

@ljvmiranda921 Thanks for the comments hahaha. Would've been searching for ages to find a solution to these things 🙈

whzup · 2018-08-17T18:05:11Z

Let's update the branch before we merge this one. I had some trouble with the rebasing of the commits.

ljvmiranda921 · 2018-08-18T07:44:07Z

tests/utils/decorators/test_decorators.py

+# Import modules
+import pytest
+import numpy as np
+import sys


I wonder where you'll use import sys for... 👨‍🚀

whzup · 2018-08-18T09:55:36Z

Ok, I guess we're ready for the merge. I think we can add change the representation of cost functions in another PR. What do you think?

ljvmiranda921 · 2018-08-18T11:28:37Z

Hi @whzup

I think we can add change the representation of cost functions in another PR. What do you think?

You mean the single_obj cost functions? I think there's no need to change them since they're all vectorized already.

Ok, I guess we're ready for the merge.

OK let me check first then I'll squash-merge this 👍

ljvmiranda921

Thank you for this @whzup ! Almost there! Just a few more things:

Please update docs/api/_pyswarms.utils.rst/. Add pyswarms.utils.decorators at the top (and since you're there, rearrange them alphabetically--decorators, functions, plotters, search)
Just a comment on __init__. The previous version is already OK. Just import cost explicitly for now. I predict this feature will be widely-used compared to other future decorators. It's better to load other decorators via pyswarms.utils.decorators

ljvmiranda921 · 2018-08-18T11:30:40Z

pyswarms/__init__.py

@@ -16,5 +16,8 @@

 from .single import global_best, local_best, general_optimizer
 from .discrete import binary
+from .utils.decorators import *
+from . import utils


Hmmm, I think the previous version was enough where cost is defined explicitly.
I think the @constraint decorator in the future is better loaded whenever the user needs it.

ljvmiranda921 · 2018-08-18T11:31:50Z

pyswarms/utils/decorators/decorators.py

+import numpy as np
+
+
+def cost(cost_func):


💯 This is good already

ljvmiranda921 · 2018-08-18T11:32:20Z

tests/utils/decorators/test_decorators.py

+    "objective_func",
+    [np.sum, np.prod]
+)
+def test_cost_decorator(objective_func, particles):


The tests are good already 💯

whzup · 2018-08-18T14:22:36Z

Hmm, I see your point. Although, what do you think about the numpy approach? Almost everything in their package can be accessed by using np.<name>. I mean, for us it's pretty clear where everything is and the project isn't that big at the moment. But for people who aren't used to the structure of this project it might be very hard to find everything and if the project becomes bigger it might become very messy to import from all these different modules. They have to search through the docs and write all these sub-module names in their code which contributes to spaghetti code. Also, if we just import the __all__ attributes into the __init__.py file of the whole package the user can still use the sub-module names etc. it's just that it's also possible to just say pyswarms.<name>. So we'd just add another option for the users who want to have very little and lucid code.
The option with the explicit statement of the sub-modules might be useful for people who want to write their own optimisation loop and the option with pyswarms.<name> for people who just want to hack together a small optimisation script.

You mean the single_obj cost functions? I think there's no need to change them since they're all vectorized already.

No, I mean all cost functions in our examples 😄

ljvmiranda921 · 2018-08-18T14:51:28Z

Hi @whzup ,

Although, what do you think about the numpy approach? Almost everything in their package can be accessed by using np.

I thought about that before, that's why there's this presence of import pyswarms as ps that I "copied" from the numpy style. The only thing that's stopping me from doing this is that we'll have instances of this:

import pyswarms as ps

ps.single.GlobalBestPSO # Good enough
ps.utils.functions.single_obj.sphere # Pretty long
ps.utils.decorators.decorators.cost # Kinda long too..

If it is possible to do ps.sphere or ps.cost but still respect the modules themselves, then maybe we can do something like that. We can load everything right away, but I'm just wary that we might have conflicting methods or something in the future.

That's the reason why I decided to have everything be loaded whenever needed. Atleast I know that users will always come for the high-level optimizers (and I'm sure the cost decorator too), but they won't really use all objective functions, and might clutter our namespace.

Right now, the design kinda follows that of Keras, where we have components separated into modules, and we load them whenever needed. The thing about numpy is that instead of components, you have operators which you will use regularly. It would then be a hassle if numpy does something similar to ours.

Still open for this idea, but for now just explicitly declare cost, we can discuss this in the later 0.X versions (v.0.5.0) 👍

ljvmiranda921 · 2018-08-18T14:53:33Z

No, I mean all cost functions in our examples 😄

Let's open-up an issue after I merge #227. We'd probably update some of our examples later + README (more animations haha)

Added a cost function decorator which allows the user to write a cost function that serves as a model for how the cost will be computed for every one particle. The decorator is tested with a pytest file where the shape and the equality with an example function is checked. I added a note to the documentation that some numpy functions will return arrays with single values in them.

whzup · 2018-08-19T13:49:40Z

Done! 👍

ljvmiranda921

LGTM!

Resolves #228 Added a cost function decorator which allows the user to write a cost function that serves as a model for how the cost will be computed for every one particle. The decorator is tested with a pytest file where the shape and the equality with an example function is checked. I added a note to the documentation that some numpy functions will return arrays with single values in them.

whzup added enhancement Feature requests documentation Documentation improvements or fixes unit tests Test-related changes v.1.1.0 In pipeline for next version labels Aug 15, 2018

whzup assigned ljvmiranda921 and whzup Aug 15, 2018

whzup requested a review from ljvmiranda921 August 15, 2018 17:52

ljvmiranda921 reviewed Aug 17, 2018

View reviewed changes

ljvmiranda921 requested changes Aug 17, 2018

View reviewed changes

ljvmiranda921 reviewed Aug 18, 2018

View reviewed changes

whzup changed the title ~~[WIP] Add cost function decorator~~ Add cost function decorator Aug 18, 2018

ljvmiranda921 mentioned this pull request Aug 18, 2018

Profiling the Topologies #206

Closed

ljvmiranda921 requested changes Aug 18, 2018

View reviewed changes

whzup added 3 commits August 19, 2018 14:41

Merge branch 'development' into decorators

c29ce4a

Merge branch 'development' into decorators

b9202c5

ljvmiranda921 approved these changes Aug 20, 2018

View reviewed changes

ljvmiranda921 merged commit 0f7d9d1 into ljvmiranda921:development Aug 20, 2018

whzup deleted the decorators branch August 31, 2018 14:59

ljvmiranda921 mentioned this pull request Jan 29, 2019

Prepare v.0.4.0 #223

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add cost function decorator #226

Add cost function decorator #226

whzup commented Aug 15, 2018 •

edited

Loading

ljvmiranda921 Aug 17, 2018

whzup Aug 17, 2018

ljvmiranda921 commented Aug 17, 2018 •

edited

Loading

ljvmiranda921 Aug 17, 2018 •

edited

Loading

whzup Aug 17, 2018

ljvmiranda921 Aug 17, 2018

ljvmiranda921 Aug 17, 2018

ljvmiranda921 Aug 17, 2018

ljvmiranda921 Aug 17, 2018 •

edited

Loading

ljvmiranda921 Aug 17, 2018

ljvmiranda921 Aug 17, 2018

whzup Aug 17, 2018

ljvmiranda921 Aug 17, 2018

ljvmiranda921 commented Aug 17, 2018

whzup commented Aug 17, 2018

whzup commented Aug 17, 2018

ljvmiranda921 Aug 18, 2018 •

edited

Loading

whzup commented Aug 18, 2018

ljvmiranda921 commented Aug 18, 2018

ljvmiranda921 left a comment

ljvmiranda921 Aug 18, 2018

ljvmiranda921 Aug 18, 2018

ljvmiranda921 Aug 18, 2018

whzup commented Aug 18, 2018 •

edited

Loading

ljvmiranda921 commented Aug 18, 2018 •

edited

Loading

ljvmiranda921 commented Aug 18, 2018

whzup commented Aug 19, 2018

ljvmiranda921 left a comment

Add cost function decorator #226

Add cost function decorator #226

Conversation

whzup commented Aug 15, 2018 • edited Loading

Description

Related Issue

Motivation and Context

How Has This Been Tested?

Types of changes

Checklist:

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ljvmiranda921 commented Aug 17, 2018 • edited Loading

ljvmiranda921 Aug 17, 2018 • edited Loading

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

ljvmiranda921 Aug 17, 2018 • edited Loading

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

ljvmiranda921 commented Aug 17, 2018

whzup commented Aug 17, 2018

whzup commented Aug 17, 2018

ljvmiranda921 Aug 18, 2018 • edited Loading

Choose a reason for hiding this comment

whzup commented Aug 18, 2018

ljvmiranda921 commented Aug 18, 2018

ljvmiranda921 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

whzup commented Aug 18, 2018 • edited Loading

ljvmiranda921 commented Aug 18, 2018 • edited Loading

ljvmiranda921 commented Aug 18, 2018

whzup commented Aug 19, 2018

ljvmiranda921 left a comment

Choose a reason for hiding this comment

whzup commented Aug 15, 2018 •

edited

Loading

ljvmiranda921 commented Aug 17, 2018 •

edited

Loading

ljvmiranda921 Aug 17, 2018 •

edited

Loading

ljvmiranda921 Aug 17, 2018 •

edited

Loading

ljvmiranda921 Aug 18, 2018 •

edited

Loading

whzup commented Aug 18, 2018 •

edited

Loading

ljvmiranda921 commented Aug 18, 2018 •

edited

Loading