Closed SagiPolaczek closed 1 year ago
Looks good; see if it's possible to be backward compatible though.
Thanks! I thought about it a bit later and it might be possible yes. Will try that on Wednesday.
@mosheraboh
Now we support both <2.0.0
, >=2.0.0
!
Please see the two last runs that differs only by the Lightning version:
I'll do another self-review now.
To support both versions I had to delete arguments (and arguments values) that brake the backward computability such as strategy
and auto_scale_batch_size
.
Ready for CR 🕺🏼
Warning - Lightning >=2.0.0 breaks backward competability
Support Lightning 2.0.0:
References:
https://lightning.ai/pages/releases/2.0.0/ (recommended!!) https://lightning.ai/docs/pytorch/latest/upgrade/migration_guide.html / thanks to @smartdanny *_epoch_end() migration guide
Changes in Fuse:
Trainer(strategy=None)
is no longer supported. We should now use strategy="auto". see here for more info.Trainer(auto_select_gpus=...)
also got removed.Tuner
andTrainer
broke up :broken_heart: -> relevant for KNIGHT’s fuse baseline.*_epoch_end()
toon_*_epoch_end()
. see changes inrun_mnist_custom_pl_imp.py
and the relevant reference above for more info.Remarks:
Support pandas 2.0.0:
References:
https://pandas.pydata.org/docs/whatsnew/v2.0.0.html#deprecations
Changes in Fuse:
df.append() was removed
. see here . Relevant for an EHR transformer dataset. Had to switch with pd.concat().