Can you add a test that it does throw a deprecation warning with 1d input?

@amueller
Given 1D input the function thinks it is a single sample array, and the later part of it doesn't really work well given only one sample. I could get it to work properly with 4 samples.

LGTM. Now it warns that 1d isn't good, at least. Having a warning for the right number of samples is for another day.

@amueller @rvraghav93
I have kept it as it because the same approach was followed in the last PR #5152

ping @ogrisel

I don't really see the point of raising a deprecation warning before raising an exception with an obscure error message:

>>> import numpy as np
>>> from sklearn.covariance import fast_mcd
>>> fast_mcd(np.arange(100))
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
/volatile/ogrisel/code/scikit-learn/sklearn/covariance/empirical_covariance_.py:75: UserWarning: Only one sample available. You may want to reshape your data array
  warnings.warn("Only one sample available. "
Traceback (most recent call last):
  File "<ipython-input-4-6294736be2de>", line 1, in <module>
    fast_mcd(np.arange(100))
  File "/volatile/ogrisel/code/scikit-learn/sklearn/covariance/robust_covariance.py", line 487, in fast_mcd
    random_state=random_state)
  File "/volatile/ogrisel/code/scikit-learn/sklearn/covariance/robust_covariance.py", line 274, in select_candidates
    random_state=random_state))
  File "/volatile/ogrisel/code/scikit-learn/sklearn/covariance/robust_covariance.py", line 146, in _c_step
    "Singular covariance matrix. "
ValueError: Singular covariance matrix. Please check that the covariance matrix corresponding to the dataset is full rank and that MinCovDet is used with Gaussian-distributed data (or at least data drawn from a unimodal, symmetric distribution.

In my opinion it would be better to directly raise a ValueError that states that computing covariance on 1D data is invalid:

X = check_array(X, ensure_2d=False, ensure_min_samples=2)
if X.ndim == 1:
    raise ValueError('Calling fast_mcd on a 1D array is invalid.')

The ensure_min_samples=2 makes sure to also raise a an informative error message when 2D data is passed with a single sample (which is not valid either).

WDYT @amueller?

@ogrisel There are some other places right now where a 1 sample or 1 feature input doesn't make sense. Would all of them would eventually be fixed this way ? Why not change check_array to accept an optional message that will be thrown when the input has too few features or samples.

Why not change check_array to accept an optional message that will be thrown when the input has too few features or samples.

I improved the error messages in #5334 by adding the estimator name when provided:

>>> from sklearn.cluster import AgglomerativeClustering
>>> AgglomerativeClustering().fit([[1, 0, -1]])
Traceback (most recent call last):
  File "<ipython-input-7-2d36cbf780b4>", line 1, in <module>
    AgglomerativeClustering().fit([[1, 0, -1]])
  File "/volatile/ogrisel/code/scikit-learn/sklearn/cluster/hierarchical.py", line 716, in fit
    X = check_array(X, ensure_min_samples=2, estimator=self)
  File "/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py", line 403, in check_array
    context))
ValueError: Found array with 1 sample(s) (shape=(1, 3)) while a minimum of 2 is required by AgglomerativeClustering.

@ogrisel There are some other places right now where a 1 sample or 1 feature input doesn't make sense.

For estimators where 1 features does not make sense I think the current state that is improved in #5334 is fine.

For the case where ensure_min_samples=2 (or more) and ensure_2d=True I think we could raise a better ValueError directly in check_array as you suggest (using the estimator name in that message).

@ogrisel If I am not wrong the error will be raised 2 versions later right ? Is there anything more you would expect here?

@ogrisel If I am not wrong the error will be raised 2 versions later right ?

In 2 versions, 1dim input will be rejected with a stronger, more generic error message like: "estimator expect 2 dimensional array-like as input, got 1 instead".

Is there anything more you would expect here?

Can you please rebase this on top of the current master and change check_array to deal specifically with the case ensure_min_samples >= 2 and ensure_2d and X.ndim == 1 to raise a specific (temporary) ValueError with an error message stating that "estimator expects at least 2 samples provided in a 2 dimensional array-like input".

@ogrisel I have rebased, but I am not sure I understand what you mean. If I modify check_array to handle the specific case that you mentioned, won't it break other people's existing code ? And what exactly do you mean by temporary ?

@ogrisel I have rebased, but I am not sure I understand what you mean. If I modify check_array to handle the specific case that you mentioned, won't it break other people's existing code ?

This will impact only code that use check_array with ensure_min_samples >= 2. Such code will always raise an exception on 1d input albeit with obscure error message as the array elements are interpreted as features of a single sample hence the ensure_min_samples >= 2 condition does not hold.

And what exactly do you mean by temporary ?

The deprecation period spans from 0.17 to 0.19. Starting in 0.19 we will always raise a ValueError on 1d input, whatever the value of ensure_min_samples.

@ogrisel Thank you, done

@ogrisel All done, thank you for your patience.

@ogrisel All done

@vighneshbirodkar can you squash all your changes please?
@ogrisel mrg?

Also: I agree, raising a value error with a good message is better.

@amueller Done

thanks :)

Thank you very much for your patience @vighneshbirodkar ! Merging now!