Feature Store is a central repository to store the feature data used in training the ML models. After the model is deployed to MLOps, the training set and test sets must be uploaded for auditing, model recreation, audit compliance, and other regulatory and business requirements.
In addition, the training and test sets must be uploaded to the Feature Store for Data Drift Detection and Model Performance Report.
To upload the training and test sets to Feature Store:
- If you are on the Model registry page, click the icon corresponding to a model for which the data needs to be uploaded to Feature Store.
- If you are already on the model Artifact or Metadata page, you will see a Feature Store link at the top menu bar. Clicking the Feature Store icon or menu item will bring the Feature Store page. Figure 4.5 shows the Feature Store page.
Figure 4.5: Feature Store page
- On the Feature Store page:
- Dataset: select either training or test set to upload the relevant dataset
- File Type: Select the file format (json, parquet, or csv)
- Browse to the data file to your local computer and upload by clicking the Submit button.
- If you deployed model directly from Momentum to MLOps, make sure to download the training and test sets from Momentum and upload to Feature Store.
- To download the training or test sets, login to Momentum
- Go to Data Upload and Exploration section
- Select the training or test sets you used in the model
- Click the Download menu item located at the top menu bar.
- The downloaded file will be a zipped file. Unzip it. You may have multiple files within the zipped file. Make sure to upload all of the training or test files one by one.
- If you deployed a third-party model (not from Momentum), make sure to acquire the training and test sets in one of the supported file formats, and the upload to the Feature Store.
- You can also upload the training and test sets to MLOps via Restful API. The API specification, with a sample Python code, is provided on the Feature Store page. Here is a sample CURL command to upload data to Feature Store via Rest API:
curl -X PUT -H ‘Content-Type:multipart/form-data’ -H ‘Authorization: Token 1234567890987654321‘ -F ‘firstname.lastname@example.org‘ http://one.accure.ai:443/mlops/v1/upload_dataset/LSTM_Test_Model/2/default_cluster/dataset_category/file_type
Import Note: For Regression models, the training and test sets must have the target field labeled as “target” even if the model was trained with a different column name of the target field. This is likely to change in the future but for now, rename your label field before uploading to Feature Store. You do not need to change anything in your model, Just the training and test datasets.