[MRG+1] Make sparse_encode give copy on write arrays

[vighneshbirodkar]

Addresses #5956

[MechCoder]

LGTM

[GaelVaroquaux]

You should not use the global random number generator in the test. Always create a local one using:

rng = np.random.RandomState(0)

and then draw samples from this random number generator.

[vighneshbirodkar]

@GaelVaroquaux I have implemented your suggestion. Can you give this a review again ?

[vighneshbirodkar]

This failed due to an HTTP failure. @MechCoder , can you re run this ?

[vighneshbirodkar]

@GaelVaroquaux Could you give this another review ?

[vighneshbirodkar]

I added come comments in the unit test to make it more clear

[ogrisel]

I am -1 on this change. mmap_mode='c' breaks under windows with large arrays:

http://stackoverflow.com/questions/24406937/scikit-learn-joblib-bug-multiprocessing-pool-self-value-out-of-range-for-i-fo

This is why I changed the default to mmap_mode='r' in joblib. The solution is to make our sparse_encode code accept to work on readonly buffers and do the copy manually only if needed. For instance you could reuse _check_copy_and_writeable from sklearn/linear_model/least_angle.py that was precisely introduced for similar purpose.

[vighneshbirodkar]

@ogrisel How about now ?

[ogrisel]

Sorry for the slow feedback @vighneshbirodkar. It looks good to me!

[ogrisel]

I will do the conflict resolution and merge.

[ogrisel]

Closing in favor of #11346.