NNPOps 0.4

[raimis]

It is time for the next release with:

• Improve the build system to use find_package(Python) instead of find_package(PythonLibs) #81
• Fix a bug the periodic boundary conditions in ANI SymmetryFunctions #83

Especially the PBC fix (#83). Does anybody working on something and wants to include it?

Ping: @peastman @RaulPPelaez @sef43

[peastman]

Sounds good to me.

[RaulPPelaez]

I would like to test this #80 a bit more.
For this #84 I say we jump straight to torch 2.0, probably solving 1.13 in the process.
Lets make 0.4 now.

[sef43]

yep go ahead!

[raimis]

A release is created: https://github.com/openmm/NNPOps/releases/tag/v0.4

Now, let's see if conda-forge is going to pick it up automatically.

[raimis]

conda-forge/nnpops-feedstock#16

[raimis]

The package builds fail because of a tolerance issue with the new tests.

[sef43]

yeah it is failing the tolerance test (which pass on the github actions CI which run on this repo) grad_error = 0.0074 fails in the assert(grad_error < 7e-3). I can reduced the tolerance to assert(grad_error < 8e-3) or even 1e-2. These seem like reasonably loose tolerances to me, are these what we expect from float32 precision gradient calculations? @raimis @peastman

[raimis]

I'm working on that.

[raimis]

Packages are available: https://anaconda.org/conda-forge/nnpops/files?version=0.4

[raimis]

@jchodera could you tweet about the release (https://github.com/openmm/NNPOps/releases/tag/v0.4)

[jchodera]

Done! https://twitter.com/openmm_toolkit/status/1632949245316845569

[raimis]

Done!