Bugfixes to train and evaluate scripts

jstmn / ikflow

Open source implementation to the paper "IKFlow: Generating Diverse Inverse Kinematics Solutions"

https://sites.google.com/view/ikflow/home

Other

35 stars 5 forks source link

Bugfixes to train and evaluate scripts #10

Closed xyyeh closed 2 months ago

xyyeh commented 2 months ago

Add minor correction to evaluate.py
Remove GPU_IDX in train*.py. GPU_IDX was not declared previously in config.py. Set devices="auto" to allow lightning to infer the max device counts. This implies that all devices in the system will be used

jstmn commented 2 months ago

why use devices="auto" vs GPU_IDX? is GPU_IDX not being set to the right gpu?

xyyeh commented 2 months ago

GPU_IDX was not declared anywhere but being used in the training scripts

jstmn commented 2 months ago

Oops, i forgot that moved to jrl

Can you change the imports to this:

from ikflow.config import DATASET_TAG_NON_SELF_COLLIDING
from jrl.config import GPU_IDX

in train_from_checkpoint.py would you mind also changing the import to this:

from jrl.config import GPU_IDX

xyyeh commented 2 months ago

Oops, i forgot that moved to jrl

Can you change the imports to this:
from ikflow.config import DATASET_TAG_NON_SELF_COLLIDING
from jrl.config import GPU_IDX
in train_from_checkpoint.py would you mind also changing the import to this:
from jrl.config import GPU_IDX

Updated.

jstmn commented 2 months ago

lgtm! thanks