Did you just replace all float32 with float64? Some parts of sklearn work in float64, right? Or not any more? The trees do and the SVM does IIRC. So in some places we want float32.

I've changed

• all float to float64.
• all float32 in some tests to float64 (under the assumption that if it passes in float64 it should pass for float32 too right?)
• all float32 in non test code is left as such.

Are these the proper things to do??

all float32 in some tests to float64 (under the assumption that if it passes in float64 it should pass for float32 too right?)

No this is a wrong assumption. 32 bit is less precise than 64 bit and some our functions and class methods might use the dtype of the input to select the precision level of their internal datastructures. Please leave the np.float32 tests unchanged.

Thanks for clarifying that!! I'll revert in a minute :)

Note that as far as I know np.dtype(np.float) always yields np.float64 on all our supported architectures (even 32 bit Python), but it's more explicit to use np.float64 directly in our code instead of relying of potentially platform specific behaviors.

@ogrisel Have reverted all float32 --> float64 changes in tests... Could you give this a go now?

lgtm

@ogrisel could you take a look? :)

Looks good to me too. +1 for merge.

@glouppe Thanks :)