Open RainierBarrett opened 3 years ago
Right now our error of the day has to do with tensor dimensions not matching what they should be in an internal cross product operation that happens when trying to get the dihedrals in the CG mapped molecules.
For some context, are the errors you're getting general hoomd-tf errors, or are they only coming up when using a system/simulation that's generated from uli-init specifically?
it looks like this PR overlaps with #27, the diffs will be easier to parse once it is merged
this error tensor dimensions not matching what they should be in an internal cross product operation
sounds to me like a numpy/mkl/scipy/tf incompatibility issue... I bet we can fix it by pinning some particular versions in the environment.yml
Merging #31 (b717914) into master (6cf51ac) will decrease coverage by
0.91%
. The diff coverage is50.00%
.:exclamation: Current head b717914 differs from pull request most recent head fafc3b6. Consider uploading reports for the commit fafc3b6 to get more accurate results
@@ Coverage Diff @@
## master #31 +/- ##
==========================================
- Coverage 93.95% 93.04% -0.92%
==========================================
Files 6 6
Lines 480 489 +9
==========================================
+ Hits 451 455 +4
- Misses 29 34 +5
Impacted Files | Coverage Δ | |
---|---|---|
uli_init/simulate.py | 92.06% <50.00%> (-2.39%) |
:arrow_down: |
Check out this pull request on
See visual diffs & provide feedback on Jupyter Notebooks.
Powered by ReviewNB
Opening this so we can discuss TF debugging stuff. @jennyfothergill the HOOMD-TF branch I'm using is here. Right now our error of the day has to do with tensor dimensions not matching what they should be in an internal cross product operation that happens when trying to get the dihedrals in the CG mapped molecules. The script I'm working on is HTF_CG_Online_Demo.py. You will also need access to the necessary gsd and .npy files which were produced using the code in this repo:
The gsd with only one PEEK molecule was made the same way but with
n_compounds=[1]
. This was needed for finding the CG mapping in the kind of hacky way that I did it. The issue right now is cropping up in thecompute()
call in this model class, which is where the TF code is getting invoked. It callshtf.mol_dihedral()
to try to get our molecule-wise batched sets of dihedral participating beads.Happy to walk through the process and how this is set up!