2 Answers
- Newest
- Most votes
- Most comments
0
Hello Nikos,
If possible please try to build the image from official image and try with further reduced image size for deploying on serverless inference. You can exclude lines from L90-L116 in the base image to reduce the size further and use the custom built final image to deploy the serverless inference endpoint. The steps for building the image are here.
answered a month ago
-1
Hi,
See part 5 of https://tutorialsdojo.com/train-and-deploy-a-scikit-learn-model-in-amazon-sagemaker/
It explains how to deploy a model trained with scikit-lean on AWS SageMaker
Best,
Didier
Thank you, but this doesn't answer my question. I specifically asked about a serverless endpoint. I know how to create a real-time endpoint like described in the link you provided.
Relevant content
- Accepted Answer
AWS OFFICIALUpdated a year ago- AWS OFFICIALUpdated 2 months ago
- AWS OFFICIALUpdated 2 months ago
- AWS OFFICIALUpdated 2 months ago
Thanks, I will try it when I find some time. I didn't know about mlio. Shall I miss some functionality during inference if I ommit those lines?
Well I tried to build locally but ran into many errors. Shouldn't there exist a ready-to-use sklearn image for serverless deployment? Or otherwise be mentioned in the documentation that serverless endpoints are not supported with sagemaker-scikit-learn-container ?
Hello Nikos,
MLIO is package useful specially during the training. In your case for inference this might not be required.
Regarding the errors during the docker build, can you test this on linux machine if not already tried?. In my local testing I found that the build fails on mac/windows systems.
"Shouldn't there exist a ready-to-use Sklearn image for serverless deployment?" - Yes there are ready to use Sagemaker images with Sklearn, but during my testing it was found out that smaller docker images seems to be working with out any issues for Serverless inference.
If you face further issues I would suggest reaching out to AWS Support.
Hi,
thank you, indeed, I was able to deploy to a serverless endpoint using image 492215442770.dkr.ecr.eu-central-1.amazonaws.com/sagemaker-scikit-learn:1.0-1-cpu-py3. But the scikit-learn package version in this image is 1.0.2 which is too old (December 2021).