Closed Ilya-Fradlin closed 6 months ago
Thanks @Ilya-Fradlin for this pull request!
Notes:
For 1, it is interesting because it didn't lead to RuntimeError
in my machine.
For 2, I only use one step lr decay in my training though. This change makes the parameter format compatible with MultiStepLR
.
For 3, thanks for catching this typo.
Merged :)
Pull Request Description:
Changes Made:
Fix in engine.py:
engine.py
where evaluating the model led to aRuntimeError
due to a device mismatch in the linesample_feats = feats[sample_mask]
. This error occurred because the indices should be on the same device as the indexed tensor. The fix ensures proper handling of device compatibility.Modification to lr_drop Parameter:
lr_drop
parameter to accept a list of integers, addressing a discrepancy in the original formulation that expected only a single integer.Correction in train_multi_scannet40.sh:
train_multi_scannet40.sh
, without it given parameters were being ignored.Additional Notes:
sample_feats = feats[sample_mask]
assample_feats
is not being used later.