Thanks for the report. I noticed the change in the API when doing the release. I didn't notice it broke the decision function. We seem to be missing some tests there.
We'll try to fix this asap.

@GaelVaroquaux and maybe @vene could you maybe have a look on how to best resolve that?

hm maybe we should just change the decision_function implementation...

you could also just use predict instead of decision_function, right?
I really don't see why we have a decision_function on regressors....

I am confused now... I can not reproduce that in master. Was that changed back again since the release?

it works for me too. Can you give a more detailed snippet to reproduce the
problem?

thanks

OK, let me see if I can isolate an appropriate example

I put off checking this, and now I actually can't seem to reproduce it in any of my code. Looks like it might have been my mistake -- oops!

No problem. Thanks for checking :)

Sorry to resurrect an old issue, but I think this bug is still lurking. Here's a minimal repro on current master:

import numpy as np
from sklearn.linear_model import LassoLars

X = np.random.random((50, 3))
y = np.random.random((50, 2))

foo = LassoLars(fit_intercept=False)
foo.fit(X, y)
foo.predict(X)

Running it results in the same problem with trying to transpose a list:

Traceback (most recent call last):
  File "demo_bug.py", line 9, in <module>
    foo.predict(X)
  File "/usr/local/lib/python2.7/dist-packages/sklearn/linear_model/base.py", line 252, in predict
    return self._decision_function(X)
  File "/usr/local/lib/python2.7/dist-packages/sklearn/linear_model/base.py", line 236, in _decision_function
    return safe_sparse_dot(X, self.coef_.T,
AttributeError: 'list' object has no attribute 'T'

Some testing shows that the bug is only triggered with a 2D y and fit_intercept=False. Changing either of those conditions makes everything work as expected.