Uniform default precision for all backends?

[matthewfeickert]

Question

As seen in the example for pyhf.set_backend the default precision for the numpy backend is 64b while the default precision for TensorFlow and PyTorch is 32b.

pyhf/src/pyhf/tensor/numpy_backend.py

Line 44 in 98bb222

self.precision = kwargs.get('precision', '64b')

pyhf/src/pyhf/tensor/jax_backend.py

Line 58 in 98bb222

self.precision = kwargs.get('precision', '64b')

pyhf/src/pyhf/tensor/tensorflow_backend.py

Line 16 in 98bb222

self.precision = kwargs.get('precision', '32b')

pyhf/src/pyhf/tensor/pytorch_backend.py

Line 18 in 98bb222

self.precision = kwargs.get('precision', '32b')

Given that there are known differences between 64b and 32b it is probably worth reconsidering / discussing if we should have a default precision of 64b for all backends. My guess is that it is non-intuitive that switching from NumPy to Torch changes the precision.

This is somewhat related to Issue #981 and was put back into my mind by @dguest and @nhartman94 (c.f. Discussion #1397) (comments welcome from them both).

[matthewfeickert]

If @lukasheinrich and @kratsg agree here I have a branch with the changes ready for a PR (after PR #1399 goes in).

[lukasheinrich]

yeah I agree!

[kratsg]

I agree. I think we had 32b at first since that was nominally the primary support when we first started with these backends. I see no reason why we shouldn't harmonize the defaults.

[matthewfeickert]

I agree. I think we had 32b at first since that was nominally the primary support when we first started with these backends.

Yeah this was my memory too.