failed doctests when PIL is not available

[fabianp]

here is the backtrace:

FAIL: Doctest: scikits.learn.datasets.base.load_sample_image
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/fabian/lib/python2.7/doctest.py", line 2153, in runTest
    raise self.failureException(self.format_failure(new.getvalue()))
AssertionError: Failed doctest test for scikits.learn.datasets.base.load_sample_image
  File "/home/fabian/dev/scikit-learn/scikits/learn/datasets/base.py", line 403, in load_sample_image

----------------------------------------------------------------------
File "/home/fabian/dev/scikit-learn/scikits/learn/datasets/base.py", line 406, in scikits.learn.datasets.base.load_sample_image
Failed example:
    china = load_sample_image('china.jpg')
Exception raised:
    Traceback (most recent call last):
      File "/home/fabian/lib/python2.7/doctest.py", line 1248, in __run
        compileflags, 1) in test.globs
      File "<doctest scikits.learn.datasets.base.load_sample_image[0]>", line 1, in <module>
        china = load_sample_image('china.jpg')
      File "/home/fabian/dev/scikit-learn/scikits/learn/datasets/base.py", line 418, in load_sample_image
        images = load_sample_images()
      File "/home/fabian/dev/scikit-learn/scikits/learn/datasets/base.py", line 389, in load_sample_images
        raise ImportError("The Python Imaging Library (PIL)"
    ImportError: The Python Imaging Library (PIL)is required to load data from jpeg files

[ogrisel]

Good catch. We need centralize the PIL related stuff somewhere and use it to disable tests on dependent methods.

[fabianp]

Was this fixed ?

[ogrisel]

I don't think so, I completely forgot about it. I won't have time today. Maybe tomorrow, unless someone else does it first.

[fabianp]

I had to comment the doctest for the release, but this still should be fixed

[amueller]

How about using PNG instead of jpeg and use pylab.imread?
Then we would never have to rely on PIL.

[amueller]

for the record: pylab is not an official requirement of sklearn, so no!

[amueller]

as an afterthought: even if we sort out the pil issue, this will crash when pylab is unavailable....

@ogrisel: Where is pil used except in this file? Or do you mean to centralize for datasets.base and datasets.test.base ? The issue of not being able to skip doctests (see issue #342) would still cause some headache.

My suggestion would be to either add skiptests as I did in #526 or reference an example as @fabianp suggested.

[ogrisel]

Yes this code is almost duplicated in datasets.base and datasets.lfw. git grep imread to check.

[fabianp]

@ogrisel : are you working on this or can we +SKIP it?. I'd really like to fix it before the release.

Same comment for issue #358

[ogrisel]

What's weird is that I did install a build of scikit-learn in virtualenv with --no-site-packages some time ago and did not experience the issue when running make. Need to retry on a recent build to check.

[amueller]

@fabianp I am +1 for SKIP (for both). What are the release plans btw?

[fabianp]

I hope to have fixes most of the issues by the end of the week, then we can think of branching and release.

[fabianp]

My priority right now is fixing the sklearn.test() failures.

[amueller]

If I can help you, just let me know.
I noticed there are some open issues for 0.10.
Do you think it would be helpful if I look into #7, #91 and/or #339?

The milestone 0.11 is due in 3 month. Is this already decided (just curious)?

By the way, I would really like your opinion on the SVM restructuring. Are you very busy at the moment?

[amueller]

My priority right now is fixing the sklearn.test() failures.

I didn't think there where any left... do you mean the warnings or the inplace import?

hmmm... I get some "AttributeError _" errors that I don't understand in cleaning up... forgot about those....

[fabianp]

Hi Andreas,

I'm OK with the SVM restructuring. I'd however postpone the merging of that for the next release instead of doing it in a rush.

It's really helpful if you can look into the issues you mentioned. Please ask if it's not clear, I realize reading them now that some of them are rather cryptic. All of them are important but not blocking so I moved them to milestone 0.11.

[amueller]

Taking time for the SVM restructuring seems sensible.
I tried to get as many small fixes in as I can before the release. I'll see if I can address any more.

[fabianp]

Great. As for the milestone, I just set it for 0.10 + 3 months because that's what our release cycle has approximately been like, but there has been no formal discussion.

[fabianp]

"Fixed"" in dfc7c5c

[ogrisel]

Thanks @fabianp.