I am trying to create a pipeline for training an image dataset via Sagemaker Pipelines. Based on the examples I understood that for all distinct stages like data preparation, model training, evaluation, and deployment there are different associated jobs determining different stages of the pipeline like Processing jobs, training jobs, conditional execution steps, fail execution steps etc. Now my question is related to the conditional execution step, it is for deciding if the created model artifact should be deployed or not based on the metric threshold set in the condition. As soon as my model metric fail to reach the desired threshold I wish to retrain my model either on some different architecture or retune my model with some Hyperparameters to adjust it's performance. This will be based on my metric score. For example my metric threshold is 0.75,and my model give me metric value of 0.8 and on retraining my model with new data I found metric value to 0.6, in that case I wish to do some parameter tuning on the existing model first. Now if I got the metric score below 0.5 would be interested in changing the architecture itself. So here the retraining step is dependent on metric value. If the metric value is too low I wish to change the model architecture itself or else I would be interested in tuning the params. How can I achieve this?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.