Open sethmlarson opened 4 years ago
@sethmlarson - this is a great area to add some really useful APIs. Maybe we should think about not just model import/export but a sub-area of features around model management.
For instance, it would be really useful to:
@tveasey may have more as well
We will have some of this management in Kibana over time, but performing low-level admin tasks and deep detail into model structure and debug would be really useful in eland.
Also, I do like storing as compressed JSON not pickle, but we should also think about the MLModel APIs being familiar to sklearn users.
This is a great initiative.
tveasey may have more as well
One thing that springs to mind is that we have a lot of training information now written into the cluster: validation loss curves, hyperparameters chosen in each optimisation round, round duration, etc. We plan to expose this information in kibana dashboard, but I also think it makes sense to be available in say a notebook if that's your preferred environment.
We should also have a think about exposing quality measures for different models as well: this likely needs additional work on the ES side, but is a longer term vision we have around workflow.
Useful when you've trained your model on a development cluster and want to export and reimport into a production cluster.
Was thinking a
to_json()
which takes abuf
and has an argumentcompress
which gzip-compresses? Then we can haveMLModel.from_json()
as the import function.cc @Winterflower @stevedodson for additional ideas here