Use log_artifact in notebook.#194
Conversation
daavoo
left a comment
daavoo
left a comment
There was a problem hiding this comment.
Actually found a bug in experiment refs in Studio (doesn't display the details of the .dvc file inside the exp ref)
| cp $HERE/code/params.yaml . | ||
| sed -e "s/base_lr: 0.01/base_lr: $BEST_EXP_BASE_LR/" -i".bkp" params.yaml | ||
| rm params.yaml.bkp | ||
| dvc remove models/model.pkl.dvc |
There was a problem hiding this comment.
not sure I understand the logic here, could you describe the workflow? do we transition from dvc add to log_artifact at some stage?
There was a problem hiding this comment.
We transition from log_artifact which is dvc add to stage output in the dvc.yaml.
There was a problem hiding this comment.
got it ... and what it was before the log_artifact where and how did we save the model?
There was a problem hiding this comment.
what it was before the log_artifact where and how did we save the model?
We didn't save it at all during the notebook state, that was the motivation for the P.R.
daavoo
deleted the
171-experiments-include-model-in-the-experimetns-inside-notebook
branch
dberenbaum
left a comment
dberenbaum
left a comment
There was a problem hiding this comment.
Thanks @daavoo! We should also add log_artifact to train.py, but it's not a blocker if we want to wait until we are ready to add type=model or other MR stuff since it should be a noop now. Do we have that tracked somewhere?
Minor non-blocking thought: the DVCLive code is starting to look a little complex. Should we add some comments to explain what's happening and make clear that you can pick and choose which calls you need?
Um. So the model would not be a stage output? |
|
No, the model can still be a stage output, that's why it would be a noop now and I think it's okay to not include it yet. Once |
3 participants
Closes #171
It is currently deployed: