WIP Allow downcast in metropolis and nuts functions.

[twiecki]

Alternative to #1253, related to #1246.

Unfortunately, the same error is also thrown if downcast is set to true:
TypeError: expected type_num 11 (NPY_FLOAT32) got 12 Apply node that caused the error: Elemwise{mul,no_inplace}(TensorConstant{(1, 1, 1, 1) of -2.0}, Reshape{4}.0)

@fhuszar Any ideas?

[twiecki]

I bet the issue is the Shared variables we're using in ArrayStepShared. That's what @nouiz meant that it would be impossible to auto-cast those.

[nouiz]

Just to clarify about the casting. If you call theano.shared() with a numpy ndarray, we will always respect its dtype. If you want to change it, call numpy.asarray(x, dtype=...) on it.

The downcast if for python int/float and with a flag, in the parameter passed to a compiled Theano function as input.

[twiecki]

Closing in favor of #1253.