(AWS WorkShop)Machine Learning Operation for Incremental Training
https://github.com/catwhiskers/mlops_incremental_learning
- Training
- retraining — Incremental training
why?
- Model drift : Differences between training and testing data
- Robustness : People get affected by ML models will deliberately alter their response
- Ground truth not available during training time : User behaviors are not predictable
Problems
- Labeling tools maintenance
- Passing human labeled results around manually
- Triggering retraining manually
- How to mange models
- Updating endpoints without down time
like NLP have many mission to labeling like content or positive / negative
This workshop will create this arch
- Training
- Deploy to Endpoint
- Submit a A2I augmented AI workflow
- Retraining
Switch to SageMaker and create notebook instance
cd ~/SageMaker/git clone https://github.com/catwhiskers/mlops_incremental_learning.git
you can see data in Jupyter
then go to SageMaker Studio and create one
then go to Ground Truth > Labeling workforces
- Amazon Mechanical Turk: price to some body on Mechanical Turk
- Private: for employee
- Vendor: for profession vendor
execute Jupyter 02-a2i-object-detection-and-retraining.ipynb file
and do with file code
you can see pipeline on Amazon SageMaker Studio
and open 03-prepare-lambda-functions.ipynb on Jupyter
and run code on Jupyter will create Lambda, you can see Lambda code in your AWS lambda
and will create SQS, you can see in your AWS SQS
need to follow Jupyter guide to set up IAM role