ensure that estimators only add private attributes and attributes with
trailing _
in cases when existing estimators don't follow this new rule, we deprecate the
attributes and make them follow this rule
* remove stuff to be removed 0.19
* more changes
* remove classes from 0.19 whatsnew
* remove _LearntSelectorMixin
* remove ProjectedGradientNMF, load_lwf_*
* minor fixes
* remove more copy from logistic regression path
* remove lda, qda from __init__.__all__
* remove pg solver in nmf from tests etc
* remove class_weight="auto" from tests
* doctest change for decision_function_shape="ovr"
* remove transfrom from tree test, minor fixes to tree tests
* some fixes in the tests
* undo changes in functions which still allow 1d input...
* also allow 1d in scale
* more test fixes...
* last test fixes in forest and tree
* svm default value change doctest failures
* pep8
* remove more class_weight="auto" stuff
* minor cosmetics in docstrings deprecated / removed behavior.
* say that store_covariance has been moved to __init__ in discriminant_analysis
* addressed comments in the PR about parameters in check_array
* update the test case for the evaluation of estimators with pandas series
* bug fix, need to check for *not* None explicitly
* updated with isinstance check if the documentation says there is acceptance of floats
* ran pep8 linter on modified files
* moving the test case to estimators_check
* add a predict function into the testing pandas.Series class
* avoid running anything beyond the newly added meta checks
* check if pandas is installed before running the specific test
* changed the order of the try-catch to check for sample_weight param beforehand
* pass on import error rather than printing something to std out
* improve test case naming and pd.Series check in the bad estimator class
* address a pep8 linter error with unused import
* pep8 warning disabled for potential unused import
* throw a warning when SkipTest is raised
* add a SkipTestWarning
* updated the whats_new.rst with this issue
* rebase and fix a spacing issue
* Remove Python 2.6 support
Some details about some slightly orthogonal changes:
* Note about cheking safely for nan is likely not valid any more (commit
introducing it is c80ca91b)
* scipy.linalg.qr econ parameter removed since scipy 0.9 in favour of
mode='economic'
* Remove unnecessary libgfortran in conda create command
* Putative fix by setting the random seed
* Revert unintended change
* Reinstate previous logic for checking for NaNs
* Reinstate change in error message
Error messages from Python 2.7 assertRegexp does not contain the
function name, in contrast with Python 3 assertRegex
* Add test for __dict__ for estimator checks
check that "predict", "transform", "decision_function" or
"predict_proba" methods do not change the state of __dict__ of any
estimator
see #7297
* Add a test to test_check_estimator
that shows that check_estimator fails on an estimator that violates this
* Fixed bug where NMF's n_iter_ attribute was set by calls to transform
* Norm inconsistency between RFE and SelectFromModel (was _LearntSelectorMixin) #2121
* safe_pwr utility
* Norm fix
* Removed safe_pwr
* 1D arrays support for norm fix
* Test case for 2d coef in SelectFromModel
* Fix numpy version requirement for norm fix
* Implement fixes suggested by @jnothman
* Add numpy version requiring the fix.
* FIX unstable cumsum in utils.random
* equal_nan = true for isclose
since numpy < 1.9 sum is as unstable as cumsum, fallback to np.cumsum
* added axis parameter to stable_cumsum
* FIX unstable sumsum in ensemble.weight_boosting and utils.stats
* FIX axis problem in stable_cumsum
* FIX unstable cumsum in mixture.gmm and mixture.dpgmm
* FIX unstable cumsum in cluster.k_means_, decomposition.pca, and manifold.locally_linear
* FIX unstable sumsum in dataset.samples_generator
* added docstring for parameter axis of stable_cumsum
* added comment for why fall back to np.cumsum when np version < 1.9
* remove unneeded stable_cumsum
* added stable_cumsum's axis testing
* FIX numpy docstring for make_sparse_spd_matrix
* change stable_cumsum from error to warning
* FIX Subclass a new MaskedArray which allows pickling even when dype=object
* TST unpickling too
* FIX Use MaskedArray from utils.fixes rather than from numpy
* FIX imports
* Don't assign a variable
* FIX np --> numpy
* Use tostring instead of tobytes for old numpy
* COSMIT pickle-able --> picklable
* use #noqa comment to turn off flake8
* TST/ENH Check if the pickled est's predict matches with the original one's
* use less nose.tools import to simplify future transition to activly developing test suites/runners
* assert_equal -> assert_array_equal in test_feature_hasher_pairs_with_string_values
and one missed ImportError that should be replaced with AttributeError
* test for py2.6 compat with except AttributeError
* fix importing of SkipTest
* force using nose in python2.6 for now
* there was no assert_dict_equal in py2.6. but we can use assert_equal
although failed test will look a little bit ugly
* remove nose imports from doc/datasets
`v0 = random_state.rand(M.shape[0])` leads to an initial residual vector in ARPACK which is all positive. However, this is not the absolute or squared residual, but a true difference. Thus, it is better to initialize with `v0=random_state.uniform(-1, 1, M.shape[0])` to have an equally distributed sign. This is the way that ARPACK initializes the residuals.
The effect of the previous initialization is that eigsh frequently does not converge to the correct eigenvalues, e.g. negative eigenvalues for s.p.d. matrix, which leads to an incorrect null-space.
- initialized all occurences of sklearn.utils.arpack.eigsh the same way it would be initialzed by ARPACK
- regression test to test behavior of new initialization
Naoya Kanai