Some TODOs for joblib 0.12.2

[qinhanmin2014]

We introduce joblib 0.12.2 in #11741, but there're still some TODOs (open an issue for further discussion)
(1) Circle CI is failing

Unexpected failing examples:
/home/circleci/project/examples/cluster/plot_feature_agglomeration_vs_univariate_selection.py failed leaving traceback:
Traceback (most recent call last):
  File "/home/circleci/project/examples/cluster/plot_feature_agglomeration_vs_univariate_selection.py", line 75, in <module>
    clf.fit(X, y)  # set the best parameters
  File "/home/circleci/project/sklearn/model_selection/_search.py", line 663, in fit
    cv.split(X, y, groups)))
  File "/home/circleci/project/sklearn/externals/joblib/parallel.py", line 981, in __call__
    if self.dispatch_one_batch(iterator):
  File "/home/circleci/project/sklearn/externals/joblib/parallel.py", line 818, in dispatch_one_batch
    self._pickle_cache)
  File "/home/circleci/project/sklearn/externals/joblib/parallel.py", line 253, in __init__
    self.items = list(iterator_slice)
  File "/home/circleci/project/sklearn/model_selection/_search.py", line 662, in <genexpr>
    for parameters, (train, test) in product(candidate_params,
  File "/home/circleci/project/sklearn/base.py", line 62, in clone
    new_object_params[name] = clone(param, safe=False)
  File "/home/circleci/project/sklearn/base.py", line 50, in clone
    return estimator_type([clone(e, safe=safe) for e in estimator])
  File "/home/circleci/project/sklearn/base.py", line 50, in <listcomp>
    return estimator_type([clone(e, safe=safe) for e in estimator])
  File "/home/circleci/project/sklearn/base.py", line 50, in clone
    return estimator_type([clone(e, safe=safe) for e in estimator])
  File "/home/circleci/project/sklearn/base.py", line 50, in <listcomp>
    return estimator_type([clone(e, safe=safe) for e in estimator])
  File "/home/circleci/project/sklearn/base.py", line 62, in clone
    new_object_params[name] = clone(param, safe=False)
  File "/home/circleci/project/sklearn/base.py", line 53, in clone
    return copy.deepcopy(estimator)
  File "/home/circleci/miniconda/envs/testenv/lib/python3.6/copy.py", line 180, in deepcopy
    y = _reconstruct(x, memo, *rv)
  File "/home/circleci/miniconda/envs/testenv/lib/python3.6/copy.py", line 274, in _reconstruct
    y = func(*args)
  File "/home/circleci/project/sklearn/externals/joblib/memory.py", line 830, in __init__
    location, cachedir))
ValueError: You set both "location='/tmp/tmp1v2mgts1' and "cachedir=False". 'cachedir' has been deprecated in version 0.12 and will be removed in version 0.14.
Please only set "location='/tmp/tmp1v2mgts1'"


/home/circleci/project/examples/compose/plot_compare_reduction.py failed leaving traceback:
Traceback (most recent call last):
  File "/home/circleci/project/examples/compose/plot_compare_reduction.py", line 119, in <module>
    grid.fit(digits.data, digits.target)
  File "/home/circleci/project/sklearn/model_selection/_search.py", line 663, in fit
    cv.split(X, y, groups)))
  File "/home/circleci/project/sklearn/externals/joblib/parallel.py", line 981, in __call__
    if self.dispatch_one_batch(iterator):
  File "/home/circleci/project/sklearn/externals/joblib/parallel.py", line 818, in dispatch_one_batch
    self._pickle_cache)
  File "/home/circleci/project/sklearn/externals/joblib/parallel.py", line 253, in __init__
    self.items = list(iterator_slice)
  File "/home/circleci/project/sklearn/model_selection/_search.py", line 662, in <genexpr>
    for parameters, (train, test) in product(candidate_params,
  File "/home/circleci/project/sklearn/base.py", line 62, in clone
    new_object_params[name] = clone(param, safe=False)
  File "/home/circleci/project/sklearn/base.py", line 53, in clone
    return copy.deepcopy(estimator)
  File "/home/circleci/miniconda/envs/testenv/lib/python3.6/copy.py", line 180, in deepcopy
    y = _reconstruct(x, memo, *rv)
  File "/home/circleci/miniconda/envs/testenv/lib/python3.6/copy.py", line 274, in _reconstruct
    y = func(*args)
  File "/home/circleci/project/sklearn/externals/joblib/memory.py", line 830, in __init__
    location, cachedir))
ValueError: You set both "location='/tmp/tmpa408j1cz' and "cachedir=False". 'cachedir' has been deprecated in version 0.12 and will be removed in version 0.14.
Please only set "location='/tmp/tmpa408j1cz'"

(2) We get many extra warnings in the examples
e.g., UserWarning: 'n_jobs' > 1 does not have any effect when 'solver' is set to 'liblinear'. Got 'n_jobs' = None.
(3) In the doc, we still have things like n_jobs : int, optional (default=1)
(4) Maybe add some explanations about the difference between n_jobs=1 and n_jobs=None?

[jnothman]

I think we need to remove all uses of cachedir...

[rth]

(3) In the doc, we still have things like n_jobs : int, optional (default=1)
(4) Maybe add some explanations about the difference between n_jobs=1 and n_jobs=None?

I think the relevant joblib lines are,

https://github.com/joblib/joblib/blob/7a17c36c83326503762a9b23d02407652ece8e3f/joblib/parallel.py#L655-L663

so this would still use n_jobs=1 by default but, I suppose, allow better interaction with a custom parallel backend (i.e. dask distributed). We should include some docs about it (possibly pointing to joblib docs?).

I think leaving it to 1 in docs might make sense: it is consistent with how we e.g. deprecate/change parameters where the value in the doc (actual value) is different from that in the code (changed to illustrate the future warning etc).

We might still might want to do some API change entry about this.

[jnothman]

I don't think we should worry too much about the default thing, at least for now. However, it might be worth documenting the n_jobs=None semantics in the Glossary entry on n_jobs.​

[qinhanmin2014]

So the final decision for (3) and (4) is to update the glossary and keep the doc? I won't argue about the decision but honestly I don't like it. Users will see n_jobs=None at the top of the doc and n_jobs : int, optional (default=1) in parameters section, which might be confused.

[jnothman]

At this stage, I'm a bit ambivalent.

…

On 14 August 2018 at 13:38, Hanmin Qin ***@***.***> wrote: So the final decision for (3) and (4) is to update the glossary and keep the doc? I won't argue about the decision but honestly I don't like it. Users will see n_jobs=None at the top of the doc and n_jobs : int, optional (default=1) in parameters section, which might be confused. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#11771 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz68AO8tTW_9OsPWICb1zgJfjfhydUks5uQkYygaJpZM4VyaQg> .

[rth]

So the final decision for (3) and (4) is to update the glossary and keep the doc? I won't argue about the decision but honestly I don't like it. Users will see n_jobs=None at the top of the doc and n_jobs : int, optional (default=1) in parameters section, which might be confused.

Well as far as I see the other alternatives would be,

• set n_jobs=None also in the docs, and specify that it corresponds to 1 job in each docstring. We can do it, but it's a lot of changes, and I wonder if it's really worth the effort, particularly now that we are a bit resource limited for the release/RC.
• revert n_jobs=1 in the code and the docs, but that would not work with custom joblib backends (initial motivation for this change), so I don't think it's a solution.

[qinhanmin2014]

revert n_jobs=1 in the code and the docs, but that would not work with custom joblib backends (initial motivation for this change), so I don't think it's a solution.

+1

set n_jobs=None also in the docs, and specify that it corresponds to 1 job in each docstring.

We've updated the glossary, so we can refer to the glossary like what we've done for random_state.

I prefer to update the doc accordingly, but I agree that this should not block the release.

[qinhanmin2014]

Closing given joblib 0.12.3.