Add MESMER-M integration tests #501
Conversation
for more information, see https://pre-commit.ci
|
Thanks - can you save the coeffs as netCDF and not pickle (I know it's a bit more annoying but we want to move away from pickle) |
veni-vidi-vici-dormivi
force-pushed
the
m_integrationtests
branch
from
17afff8
to
f97ea6e
Compare
for more information, see https://pre-commit.ci
|
I haven't looked into it so not sure what is going on but a bit strange that none of the environments get it right. I thought this does not have to be high priority but then it would be good to stabilize the numerics for memser-m. What helped for the covariance thingy was testing the code on cfc vs exo, which have different cpus with different sets of SIMD instructions. But I am not really working today - if I find time I'll take a look tomorrow or else after our retreat... |
|
Do you happen to know if the estimates are stable for each OS? Could we do os-dependent tolerances? (Not asking you to sink more time into this PR atm...) |
Sorry for never replying to this, yes the estimates are different from OS to OS, I remember I needed to increase the tolerance somewhere for windows to also pass. I guess we could do OS dependent tolerances but it's kind of awkward, since the tolerance then depends on which OS you produced the data with originally. Like now with the data from macOS, windows would need a larger tolerance, but if we ever push a dataset that was generated with windows it would have to be the other way around... so I'm not sure. but I like #563, although it is going back to what we did in #402, but oh well. |
Good point! Let's not do this. |
|
|
|
Yes I agree that that's not so nice and I also am not that convinced of having one parameter file per ESM anymore, so it's fine for me to have a file or netcdf group? 👀👀 for each statistical method. However, I still think it is good to have the lambda parameters also along a coeff dimension in case we or someone else ever implements different lambda functions. |
|
Ah okay we had that already, just without coordinates. Yeah I would keep the coordinates but we can revert the unique naming |
CHANGELOG.rst