[MRG] Have Travis actually execute doctests

[rtavenar]

Hi there,

As it is coded at the moment, doctests are not checked by Travis, since the call in test_ot does not raise any error. This PR is an attempt to fix that.

[rtavenar]

At the moment, I have:

• moved GPU doctests to test_gpu (+added an ignore flag on the ot/gpu folder) since otherwise doctest would fail on importing the module if GPU is not supported
• edited travis.yml file so that doctests are executed
• fixed obvious bugs in doctests

I have not fixed bugs related to numpy array formatting, let me know which version of formatting I should use or if there is a clever way to deal with those in doctests.

Also, there are the following tests that do not pass and that would need checking:

284         log dictionary return only if log==True in parameters
285 
286     Examples
287     --------
288 
289     >>> import ot
290     >>> a=[.5, .15]
291     >>> b=[.5, .5]
292     >>> M=[[0., 1.],[1., 0.]]
293     >>> ot.sinkhorn_knopp_unbalanced(a, b, M, 1., 1.)
UNEXPECTED EXCEPTION: AttributeError("module 'ot' has no attribute 'sinkhorn_knopp_unbalanced'")
Traceback (most recent call last):

  File "/usr/local/Cellar/python/3.7.3/Frameworks/Python.framework/Versions/3.7/lib/python3.7/doctest.py", line 1329, in __run
    compileflags, 1), test.globs)

  File "<doctest ot.unbalanced.sinkhorn_knopp_unbalanced[4]>", line 1, in <module>

AttributeError: module 'ot' has no attribute 'sinkhorn_knopp_unbalanced'

Should I add this as a member in __init__.py ?

228     >>> numItermax = 300000
229     >>> a = ot.utils.unif(n_source)
230     >>> b = ot.utils.unif(n_target)
231     >>> rng = np.random.RandomState(0)
232     >>> X_source = rng.randn(n_source, 2)
233     >>> Y_target = rng.randn(n_target, 2)
234     >>> M = ot.dist(X_source, Y_target)
235     >>> method = "ASGD"
236     >>> asgd_pi = ot.stochastic.solve_semi_dual_entropic(a, b, M, reg, method, numItermax)
237     >>> print(asgd_pi)
Expected nothing
Got:
    [[2.55489470e-02 9.96268666e-02 1.76769854e-02 4.34379685e-06]
     [1.21589227e-01 1.25320316e-02 1.30343724e-03 7.43244699e-03]
     [3.55891217e-03 7.61047795e-02 6.31933164e-02 1.34770804e-07]
     [2.61300196e-02 3.34012346e-02 8.29155469e-02 4.10341751e-04]
     [9.79012843e-03 7.47427885e-04 1.07693028e-02 1.21550284e-01]
     [2.15528997e-02 8.98205807e-04 1.86239266e-03 1.18543645e-01]
     [4.15119582e-02 2.65801363e-02 7.23542541e-02 2.41079424e-03]]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:237: DocTestFailure
______________________________________________________________________________________________________________ [doctest] ot.stochastic.batch_grad_dual ______________________________________________________________________________________________________________
510     >>> batch_size = 3
511     >>> log = True
512     >>> a = ot.utils.unif(n_source)
513     >>> b = ot.utils.unif(n_target)
514     >>> rng = np.random.RandomState(0)
515     >>> X_source = rng.randn(n_source, 2)
516     >>> Y_target = rng.randn(n_target, 2)
517     >>> M = ot.dist(X_source, Y_target)
518     >>> sgd_dual_pi, log = ot.stochastic.solve_dual_entropic(a, b, M, reg, batch_size, numItermax, lr, log)
519     >>> print(log['alpha'], log['beta'])
Expected nothing
Got:
    [0.65417474 1.28464039 0.75387716 0.15738988 0.55418466 1.06094533
     0.19524193] [0.50659034 0.58415035 1.29294164 2.27677175]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:519: DocTestFailure
___________________________________________________________________________________________________________ [doctest] ot.stochastic.c_transform_entropic ____________________________________________________________________________________________________________
307     >>> numItermax = 300000
308     >>> a = ot.utils.unif(n_source)
309     >>> b = ot.utils.unif(n_target)
310     >>> rng = np.random.RandomState(0)
311     >>> X_source = rng.randn(n_source, 2)
312     >>> Y_target = rng.randn(n_target, 2)
313     >>> M = ot.dist(X_source, Y_target)
314     >>> method = "ASGD"
315     >>> asgd_pi = ot.stochastic.solve_semi_dual_entropic(a, b, M, reg, method, numItermax)
316     >>> print(asgd_pi)
Expected nothing
Got:
    [[2.54106554e-02 9.97268805e-02 1.77153111e-02 4.29584338e-06]
     [1.21547758e-01 1.26085817e-02 1.31292434e-03 7.38787857e-03]
     [3.53484453e-03 7.60777888e-02 6.32443764e-02 1.33102110e-07]
     [2.59763391e-02 3.34190138e-02 8.30561694e-02 4.05620586e-04]
     [9.83147177e-03 7.55426727e-04 1.08972134e-02 1.21373031e-01]
     [2.16515127e-02 9.08136849e-04 1.88517429e-03 1.18412319e-01]
     [4.13068530e-02 2.66194449e-02 7.25455335e-02 2.38531153e-03]]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:316: DocTestFailure
_________________________________________________________________________________________________________ [doctest] ot.stochastic.coordinate_grad_semi_dual _________________________________________________________________________________________________________
058     >>> numItermax = 300000
059     >>> a = ot.utils.unif(n_source)
060     >>> b = ot.utils.unif(n_target)
061     >>> rng = np.random.RandomState(0)
062     >>> X_source = rng.randn(n_source, 2)
063     >>> Y_target = rng.randn(n_target, 2)
064     >>> M = ot.dist(X_source, Y_target)
065     >>> method = "ASGD"
066     >>> asgd_pi = ot.stochastic.solve_semi_dual_entropic(a, b, M, reg, method, numItermax)
067     >>> print(asgd_pi)
Expected nothing
Got:
    [[2.55691325e-02 9.95968638e-02 1.76867732e-02 4.37336996e-06]
     [1.21563085e-01 1.25156756e-02 1.30284921e-03 7.47553294e-03]
     [3.56135291e-03 7.60739345e-02 6.32217198e-02 1.35674205e-07]
     [2.61398154e-02 3.33773232e-02 8.29270403e-02 4.12964016e-04]
     [9.74052071e-03 7.42829745e-04 1.07122029e-02 1.21661590e-01]
     [2.14460994e-02 8.92780448e-04 1.85272634e-03 1.18665537e-01]
     [4.15211350e-02 2.65570239e-02 7.23531567e-02 2.42582723e-03]]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:67: DocTestFailure
__________________________________________________________________________________________________________ [doctest] ot.stochastic.sag_entropic_transport ___________________________________________________________________________________________________________
139     >>> numItermax = 300000
140     >>> a = ot.utils.unif(n_source)
141     >>> b = ot.utils.unif(n_target)
142     >>> rng = np.random.RandomState(0)
143     >>> X_source = rng.randn(n_source, 2)
144     >>> Y_target = rng.randn(n_target, 2)
145     >>> M = ot.dist(X_source, Y_target)
146     >>> method = "ASGD"
147     >>> asgd_pi = ot.stochastic.solve_semi_dual_entropic(a, b, M, reg, method, numItermax)
148     >>> print(asgd_pi)
Expected nothing
Got:
    [[2.53339520e-02 9.98586476e-02 1.76602698e-02 4.27351445e-06]
     [1.21514908e-01 1.26600440e-02 1.31245306e-03 7.36973751e-03]
     [3.52680737e-03 7.62352226e-02 6.30949803e-02 1.32509197e-07]
     [2.59514121e-02 3.35322774e-02 8.29691078e-02 4.04345580e-04]
     [9.84960357e-03 7.60114347e-04 1.09163422e-02 1.21331083e-01]
     [2.16902366e-02 9.13721224e-04 1.88837841e-03 1.18364807e-01]
     [4.12767387e-02 2.67158275e-02 7.24862142e-02 2.37836243e-03]]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:148: DocTestFailure
________________________________________________________________________________________________________ [doctest] ot.stochastic.sgd_entropic_regularization ________________________________________________________________________________________________________
599     >>> batch_size = 3
600     >>> log = True
601     >>> a = ot.utils.unif(n_source)
602     >>> b = ot.utils.unif(n_target)
603     >>> rng = np.random.RandomState(0)
604     >>> X_source = rng.randn(n_source, 2)
605     >>> Y_target = rng.randn(n_target, 2)
606     >>> M = ot.dist(X_source, Y_target)
607     >>> sgd_dual_pi, log = ot.stochastic.solve_dual_entropic(a, b, M, reg, batch_size, numItermax, lr, log)
608     >>> print(log['alpha'], log['beta'])
Expected nothing
Got:
    [0.64640722 1.26827311 0.7740418  0.15379161 0.55547374 1.04208024
     0.19900565] [0.51303392 0.58323389 1.28687355 2.25593202]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:608: DocTestFailure
____________________________________________________________________________________________________________ [doctest] ot.stochastic.solve_dual_entropic ____________________________________________________________________________________________________________
695     >>> batch_size = 3
696     >>> log = True
697     >>> a = ot.utils.unif(n_source)
698     >>> b = ot.utils.unif(n_target)
699     >>> rng = np.random.RandomState(0)
700     >>> X_source = rng.randn(n_source, 2)
701     >>> Y_target = rng.randn(n_target, 2)
702     >>> M = ot.dist(X_source, Y_target)
703     >>> sgd_dual_pi, log = ot.stochastic.solve_dual_entropic(a, b, M, reg, batch_size, numItermax, lr, log)
704     >>> print(log['alpha'], log['beta'])
Expected nothing
Got:
    [0.6459203  1.25256745 0.77168788 0.16400835 0.54905627 1.0560163
     0.20497883] [0.50990995 0.5803749  1.2790953  2.27485523]

/Users/tavenard_r/Documents/costel/src/POT/ot/stochastic.py:704: DocTestFailure
_________________________________________________________________________________________________________ [doctest] ot.stochastic.solve_semi_dual_entropic __________________________________________________________________________________________________________
401     >>> numItermax = 300000
402     >>> a = ot.utils.unif(n_source)
403     >>> b = ot.utils.unif(n_target)
404     >>> rng = np.random.RandomState(0)
405     >>> X_source = rng.randn(n_source, 2)
406     >>> Y_target = rng.randn(n_target, 2)
407     >>> M = ot.dist(X_source, Y_target)
408     >>> method = "ASGD"
409     >>> asgd_pi = ot.stochastic.solve_semi_dual_entropic(a, b, M, reg, method, numItermax)
410     >>> print(asgd_pi)
Expected nothing
Got:
    [[2.56160093e-02 9.95898692e-02 1.76468652e-02 4.39912542e-06]
     [1.21562094e-01 1.24917930e-02 1.29752011e-03 7.50573569e-03]
     [3.57141868e-03 7.61439936e-02 6.31415940e-02 1.36608488e-07]
     [2.62132630e-02 3.34075088e-02 8.28205702e-02 4.15800902e-04]
     [9.71011104e-03 7.39103615e-04 1.06351665e-02 1.21772762e-01]
     [2.13756023e-02 8.88154937e-04 1.83909771e-03 1.18754288e-01]
     [4.16190236e-02 2.65690541e-02 7.22276750e-02 2.44139012e-03]]

Could someone check if those values are reasonable?

1552 
1553     Examples
1554     --------
1555 
1556     >>> n_s = 2
1557     >>> n_t = 4
1558     >>> reg = 0.1
1559     >>> X_s = np.reshape(np.arange(n_s), (n_s, 1))
1560     >>> X_t = np.reshape(np.arange(0, n_t), (n_t, 1))
1561     >>> empirical_sinkhorn_divergence(X_s, X_t, reg)
Expected:
    array([2.99977435])
Got:
    array([1.49988718])

This test definitely fails: is it the test that is wrong or the value returned by POT?

[rtavenar]

OK, things should be (close to) OK now, with tests executed and passing.

By the way, there was a bug in unbalanced.py (Python 2) due to division by an integer. This is now fixed. Let me know what I can do to improve.

[rflamary]

This is awesome, thank you @rtavenar

I see no problem in the code and will do the merge.