Closed chaitjo closed 7 months ago
I found that the error was indeed due to the validation metrics not being computed due to a bug, which got ignored by the try-except logic here: https://github.com/microsoft/protein-frame-flow/blob/main/models/flow_module.py#L177
In particular, there's no variable named CA_IDX in metrics. (https://github.com/microsoft/protein-frame-flow/blob/main/models/flow_module.py#L179)
My work around was to import residue_constants
and use that to get ca_idx
following how its done in the rest of the code. In the import statements, add:
from data import residue_constants
And change line 179 to:
ca_idx = residue_constants.atom_order['CA']
ca_ca_metrics = metrics.calc_ca_ca_metrics(final_pos[:, ca_idx])
This seems to work and I've opened a PR if you think this is the appropriate way to fix the issue.
Thanks for the catch and fix! This looks good to me.
After making the fix in #17 and getting the training script to start running with the default configs, I'm encountering the following error wherein the code breaks after completing one training and one validation epoch. The error message suggests that the metric
valid/non_coil_percent
which is used for saving checkpoints was not logged/saved/computed.Please let me know if there's a quick work around.