The third swimlane in Fig. 1 illustrates the Training work stream. In this work stream, the ML model coded in the Modeling work stream is trained and validated; the validated trained model is registered for deployment in the Inference work stream. The activities in this work stream can be done mannually, or automated whereever possible especially for recurrent training.
|Train Model||The pre-trained ML model coded in the Modeling work stream is trainined with a updated training set using the selected hyperparameters.||Pre-trainined model & hyperparameters & training set||Trained model|
|Validate Model||The new trainied ML model is validated with a testing set on the inference accuracy. Only the model with an expected accuracy is used for deployment.||Trained model & testing set||Validated model|
|Register Model||The validated model with an expected accuracy is registered to the registry for deployment to the Inference work stream.||Validated model||Trained model published to registry|
The coded ML model and the selected hyperparameters from the Modeling work stream are passed to this activity. The pre-trainined ML model is recurrently trained when a updated training set is generated. This activity can be done mannually. If the trained model needs to be updated regularly, the training process should be programmatically automated. Note that the training set should be representative of the current data pattern for inference.
The newly trained model is validated against performing inference on a testing set. A validation rule can be set to define the criteria on the acceptance of an updated trained model. For example, the precision and recall of the updated trained model must be above predefined thresholds. Note that the testing set should be representative of the current data pattern for inference.