Should spox.opset.ai.onnx.vXX.const default to float64 for float input?

[cbourjau]

Numpy defaults to float64 when presented with python floats:

np.array([3.14]).dtype == np.dtype("float64")

However, the const convenience method defaults to float32. We should consider changing this behavior in order to stay closer to numpy's behavior.

[JakubBachurskiQC]

This is indeed a good question. We did end up using npt.DTypeLike in cast which made op.cast(x, to=float) cast to float64. I could see that being an argument for op.const(x: float) becoming float64.
It would be however unfortunate to isolate ML/DL-oriented users that would prefer float32-oriented behaviour. I'm not sure what would be the best way of approaching this.

[JakubBachurskiQC]

I would propose changing semantics of const to arg -> op.constant(value=numpy.array(arg)), or removing it entirely.
It's not a battle worth fighting further by trying to come up with conventions! Thoughts @cbourjau ?

[cbourjau]

I think having const is very nice. op.constant(value=np.array(42)) is certainly less nice to type out than op.const(42). Delegating it all to numpy, as you suggest, seems to be the easier and right thing to do, though!

[JakubBachurskiQC]

Using an equivalent of this I think the only thing that was missing was having a dtype argument, defaulting to what numpy.array does when none is provided.

[cbourjau]

I'm not sure I understand what you mean. Are you saying we should have something like op.const(arg, dtype=np.float32)?

[JakubBachurskiQC]

Yes - as otherwise op.const(3.14) is always an f64. Writing op.const(np.array(3.14, np.float32)) is a bit longer than it could be: op.const(3.14, np.float32).

[cbourjau]

Yes, that makes sense!

[JakubBachurskiQC]

I'm debating whether we should be using Constant for this at all. It seems that initialisers are the actual expected mechanism for constants on this IR version. We'll get to this issue if we want to do #25 - as other opset modules (well, only ai.onnx.ml currently) will also want to construct constants. So maybe the actual best approach is to expose a spox.const|initializer (ArrayLike) -> Var public function that uses initialisers underneath?