Emitting a RuntimeWarning would probably be a good idea here

Emitting a RuntimeWarning would probably be a good idea here

My suggestion is to convert np.nan to np.nan as with float32

np.array([np.nan]).astype(np.float32) 

outputs

array([nan], dtype=float32)

Integers can't store nan though? That's like saying np.array(2).astype(bool) should give 2.

I've just ran into this. So nan is converted tot the smallest 32bit signed integer. I don't know if this is deliberate, but this behavior coincidentally suits my needs.

However, it does not seem to be a very deliberate design, because np.int32(np.nan) throws a ValueError.

It feel it should be documented, at the least, under the form of a warning or deliberate design.

@mxt3 note that NumPy currently returns whatever (int32)NaN returns in C, and that is unfortunately "undefined", I think. IIRC, there are machines where you may get a different result unfortunately, e.g. 0 (there should be an issue about that open somewhere).

This will now give a RuntimeWarning on main (will be a while until a release with it). There is another issue about this open at gh-14412, which I will keep open as a point for discussing whether more than a warning (settable using np.errstate(invalid="raise"/"ignore") would be a nice further step.