Hi Matt.
Glad my example was useful in some way. I guess knowing exactly which
Cython routines to test for what is what I meant about where to start.
Thanks for the tip about the stream frontend.
- Casey
On Mon, Sep 24, 2012 at 1:52 PM, Matthew Turk
Hey Casey and Anthony,
Hi Anthony.
I completely agree that we should target the level of functions actually performing the projection rather than yt's organization. The mock frontend suggestion was just a hack to get there. I don't know if there's a way around it though...
Here's an example of what I sorted through to get to projections: - Load a test plotfile, check pf.h.proj to find it's source. - Read through data_objects/hierarchy.py and utilities/parallel_tools/parallel_analysis_interface.py to find where
On Mon, Sep 24, 2012 at 4:20 PM, Casey W. Stark
wrote: proj is attached, can't find it. - The proj docstring says it is a reference to AMRQuadProj. Can't find a class by that name. - Search data_objects sources for "proj", find AMRProjBase.
So it looks like the functionality is wrapped up in the __project_level and _project_grid methods. I can't think of a way to test those without creating an AMRProjBase, and that requires a staticoutput object.
You're right, the projection stuff as *projections* is not easy to test. But in terms of testing the underlying code, which is wrapped up in a Cython class called QuadTree, I think it could be done. The steps you're describing are actually all part of the existing answer testing machinery, which performs a couple things and verifies that they don't change over time:
1) Project some fields from the disk 2) Project a couple derived fields 3) Project a derived field that requires spatial derivatives 4) Project the "Ones" field, which should be 1.0 everywhere.
So these things are done, but it is also possible that the specific quadtree functionality could be tested, in isolation from the projection. I think this may be oneo f the things Anthony is talking about -- answer testing can handle the big, complex items, and by breaking down to the fundamentals we can address isolated items from a unit testing perspective.
So unfortunately, I think it would still come down to having a fake frontend. It's not ideal, but it seems like any more isolation would
require
big rewrites to yt.
One fun thing that is not usually known is that we have a fake frontend already, it just doesn't get used much. It's called the "Stream" frontend and it was designed originally to be used in ParaView, but now gets used by the (new, not-yet-documented/released) load_uniform_grid function as well as by Hyperion, the RT code by Tom R. It can set up AMR as well as static mesh. It's not terribly well documented, but there are examples on the wiki.
One thing I've been thinking about is actually creating a couple fake outputs, which could be defined analytically with spheres of overdensity inside them. In principle, if we added refinement criteria, we could make this relatively complex data that was defined with only a few lines of code, but spun up a big in-memory dataset.
(This exact thing is on my list of things to do and then to output in GDF, by the way...)
That I think could come, down the road a bit. The refinement criteria wouldn't be too bad to implement, especially since we already have the grid splitting routines. I just don't think we should focus on it at the moment. But the uniform grid creation and loading works already -- I used it this morning. You can do it with:
from yt.frontends.stream.api import load_uniform_grid ug = load_uniform_grid({"VelocityNorm":data1, "Density":data2}, [359, 359, 359], 1.0)
the list is the dimensions of the data and the value is the to-cm conversion.
Of course, I could be missing something. Matt, can you think of a better way?
I think for this specific example (and your damningly complex tracing of things through the source ...) the easiest thing to do is isolate the Cython routine, which it seems I was able to do only because I wrote it and which seems quite buried in the code, and to also provide high-level machinery for faking a frontend.
-Matt
- Casey
On Mon, Sep 24, 2012 at 11:02 AM, Anthony Scopatz
Helo Casey,
Sorry for taking the whole weekend to respond.
I would like to help with this, but it's difficult to figure out where to start.
Not to worry. I think that any of the items listed at the bottom of
Matt's
original email would be a great place to start.
Say I want to test projections. I make a fake 3D density field, maybe something as simple as np.arange(4**3).reshape((4, 4, 4)). I write
down the
answer to the x-projection. Now all I need to do is call assert_allclose(yt_result, answer, rtol=1e-15), but I don't know what
wrote: pieces
of low-level yt stuff to call to get to `yt_result`. Hopefully that's clear...
Maybe this comes down to creating a fake frontend we can attach fields to?
Actually, I disagree with this strategy, as I told Matt when we spoke last week. What is important is that we test the science and math parts of the code before, if ever, dealing with the software architecture that surrounds them.
Let's taking your example of projections. What we need to test is the actual function or method which actually slogs through the projection calculation. In many cases in yt these functions are not directly attached to the front end but live in analysis, visualization or utilities subpackages. It is these such packages that we should worry about testing. We can easily create routines to feed them sample data.
On the other hand, testing or mocking things like frontends should be a very low priority. At the end of the day what you are testing here is pulling in data from disk or other sources. Effectively, this is just re-testing functionality present in h5py, etc. That is not really our job. Yes, in a perfect world, front ends would be tested too. But I think that the priority should be placed on things like the KDTree.
Be Well Anthony
- Casey
On Fri, Sep 21, 2012 at 2:42 PM, Matthew Turk
wrote: Hi all,
As some of you have seen (at least Stephen), I filed a ticket this morning about increasing testing coverage. The other night Anthony and I met up in NYC and he had something of an "intervention" about the sufficiency of answer testing for yt; it didn't take too much
work
on his part to convince me that we should be testing not just against a gold standard, but also performing unit tests. In the past I had eschewed unit testing simply because the task of mocking data was quite tricky, and by adding tests that use smaller bits we could cover unit testable areas with answer testing.
But, this isn't really a good strategy. Let's move to having both. The testing infrastructure he recommends is the nearly-omnipresent nose:
http://nose.readthedocs.org/en/latest/
The ticket to track this is here:
https://bitbucket.org/yt_analysis/yt/issue/426/increase-unit-test-coverage
There are a couple sub-items here:
1) NumPy's nose test plugins provide a lot of necessary functionality that we have reimplemented in the answer testing utilities. I'd like to start using the numpy plugins, which include things like conditional test execution, array comparisons, "slow" tests, etc etc. 2) We can evaluate, using conditional test execution, moving to nose for answer testing. But that's not on the agenda now. 3) Writing tests for nose is super easy, and running them is too.
Just
do:
nosetest -w yt/
when in your source directory.
4) I've written a simple sample here:
https://bitbucket.org/yt_analysis/yt-3.0/src/da10ffc17f6d/yt/utilities/tests...
5) I'll handle writing up some mock data that doesn't require
shipping
lots of binary files, which can then be used for checking things that absolutely require hierarchies.
--
The way to organize tests is easy. Inside each directory with testable items create a new directory called "tests", and in here toss some scripts. You can stick a bunch of functions in those scripts.
Anyway, I'm going to start writing more of these (in the main yt repo, and this change will be grafted there as well) and I'll write back once the data mocking is ready. I'd like it if we started encouraging or even mandating simple tests (and/or answer tests) for functionality that gets added, but that's a discussion that should be held separately.
The items on the ticket:
* kD-tree for nearest neighbor * Geometric selection routines * Profiles * Projections -- underlying quadtree * Data object selection of data containers * Data object selection of points * Orientation class * Pixelization * Color maps * PNG writing
Is anyone willing to claim any additional items that they will help write unit tests for?
-Matt _______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
_______________________________________________ yt-dev mailing list yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org