morning folks, I'd like to know your tooling stack...
# general
a
morning folks, I'd like to know your tooling stack if you have built end to end ml ops pipeline for the nixtla libraries - esp ML Forecast. Our current approach is 1. to use cloud pickle to export the model binary 2. track experiments using our own custom built scripts. We are specifically looking into either mlflow or sagemaker.
m
@Kevin Kho and @José Morales: this might be a good moment to build the MLFLOW tutorial.
k
this is assigned to me 😅. I can try to get it done this week
👍 1
🙌 1
a
@Kevin Kho do you have an update on this? also, is it possible to use kubeflow for nixtla libs?
k
Hey Arun, sorry I’ve been busy I wasn’t able to get to it. I’ll do it for sure tomorrow! I think you should be able to use Kubeflow. You mean in a pipeline?
a
thanks @Kevin Kho, right, in the pipeline
k
I think you would just call Nixtla in the pipeline right?
👍 1
a
I agree, kubeflow should be fine. I look forward to the mlflow doc.
k
The PR is out here but you can also check the version here for the MLFlow integration The PR is the same, it just cleans it up a bit and simplifies
🙌 1
a
Thank you so much @Kevin Kho @Max (Nixtla)
@Kevin Kho We use the cloudpickle for the MLForecast, do you have a similar doc for MLForecast (like you have for StatsForecast)?
k
I don’t think there is…I could make one this Friday. Asking @José Morales to confirm we don’t have a doc for MLFlow + MLForecast?
j
We don't. I've also thought of adding save and load methods that remove attributes that aren't required for inference, so it's probably best to wait for those, WDYT?
k
Ah ok
a
@Kevin Kho @José Morales morning folks! would be great to have a write up for this even for the existing features, appreciate the help here!
k
Will talk to Jose about it after some items I’m working on
🙌 1