change assertEqual to assertAlmostEqual in tests

[huanzhang12]

We should use assertAlmostEqual instead of assertEqual for comparing floating point numbers in tests. Otherwise sometimes there will be some tiny differences between the results and the tests fail.
Similarly, we should use np.testing.assert_almost_equal to replace assertListEqual.
I searched assertEqual and assertListEqual in all tests and it seems this is the last one we should replace.

[guolinke]

@huanzhang12
This is used for checking the consistency of model file, so it should be exact same.

refer to #31 (comment)

[huanzhang12]

@guolinke I see. So what actually happens when the test fails?
I get this on my laptop, compiled without GPU support:

----------------------------------------------------------------------
running test_basic.py
([('training', u'auc', 0.982663878622836, True)], [('valid_1', u'auc', 0.98148148148148151, True)])
([('training', u'auc', 0.99521007586072752, True)], [('valid_1', u'auc', 0.98290598290598286, True)])
([('training', u'auc', 0.99632043052583807, True)], [('valid_1', u'auc', 0.98433048433048431, True)])
F
======================================================================
FAIL: test (__main__.TestBasic)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test_basic.py", line 50, in test
    self.assertEqual(*preds)
AssertionError: 0.028836746890177237 != 0.028836746890177251

----------------------------------------------------------------------
Ran 1 test in 0.065s

FAILED (failures=1)

[guolinke]

@huanzhang12
not very sure, maybe the sumup order by multi-threading.

https://github.com/Microsoft/LightGBM/blob/master/src/boosting/gbdt.cpp#L904-L907

you can try to remove the omp in these predict function and check it again ?

[huanzhang12]

@guolinke It is indeed the multi-threading problem in gbdt.cpp. After removing the pragma the test can pass on the two machines previously failed the test.

So what should we do here? I think we can either do assertAlmostEqual with more digits (like 15 digits), or setting num_threads=1 for this test. Which way do you prefer? Or do you have any better suggestions on this issue?

[guolinke]

I prefer set num_threads=1 in test_basic.py.

[huanzhang12]

OK, I have amended my commit to use one thread on this test.
Also I added a comment before the assertEqual test in case someone else raises the same question.