Leaving aside the training loss, the optimizer (ADAM), the number of U-NET blocks (tuned to a meaningful target receptive field for the problem) and the number of filters per layer (set to the maximum fitting in gpu memory for batch size = 1), what would be the most critical hyperparameters, and some sensible guidelines to tune them to achieve the best prediction accuracy without having to perform a multi-gpu extensive/expensive search for each problem?



Your Answer

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