Closed jflat06 closed 7 months ago
@fdimaio
I think this is good to go.
All of the wholepose/block tests are passing on all energy terms.
You can see the data format of the new baselines in tmol/tests/data/term_baselines. I formatted it in yaml like we had talked about. I can tweak the formatting if we want alterations.
I also did a small monkeypatch to torch to format the gradcheck failure output a bit better (Example):
E torch.autograd.gradcheck.GradcheckError: Difference between analytical and numerical tensors exceeds tolerances:
E index | analytical | numerical | difference
E [0 0] | -0.09107021 | 0.0 | 0.09107021
E [1 0] | 0.0971392 | -0.23841858 | 0.3355578
E [2 0] | 0.14213827 | 0.0 | 0.14213827
E [3 0] | 0.01525695 | 0.0 | 0.01525695
E [4 0] | -0.07564153 | -0.47683716 | 0.40119565
E [5 0] | -0.07978861 | -0.47683716 | 0.39704853
E [6 0] | -0.349249 | -0.47683716 | 0.12758815
E [7 0] | -0.42451367 | -0.47683716 | 0.05232349
E [8 0] | 0.2448904 | 0.0 | 0.2448904
This only shows you the failing comparisons and their index. The baseline comparisons also use this formatting.
This is almost ready to go, but is currently stuck on an issue with dunbrack block pair reweighted gradcheck failing. I haven't been able to figure this out, so it may help to get some other eyes on the code.