Issue with Loss Convergence During Training

[arenas7307979]

I have been training my GNSS network using the following setup:

num_epochs_seq = [1000, 1000, 1000, 1000, 1000] and num_epochs_seq = [100, 100, 100, 100, 100]

Despite training for a substantial number of epochs, I am unable to achieve loss convergence. No matter how many epochs I train, the loss remains stagnant.

Has anyone else experienced similar issues with their data? Is this behavior consistent with your testing results? Any insights or suggestions to resolve this would be greatly appreciated.

Thank you!

[AILocAR]

Hey there, thank you for your interest!

According to your training results,you have almost completed the training progress. But you may do better by decreasing your learning rate. Now, you can try to test your network.

I want to point out that you cannot get the perfect zero training loss that are common in classification tasks because PrNet is designed for regression problem and the labels are noisy(even though we have filtered them).

[arenas7307979]

Thank you for your insightful paper. Following your suggestions, I adjusted the learning rate and obtained the following results:

Could you kindly verify if these results are reasonable and comparable to yours? Also, would RMSprop be more effective than Adam for handling noisy data in this context? As a beginner, I appreciate your patience with any basic questions I might have.

Best regards, arenas.

[AILocAR]

I think your results are good. You can try to plot out your trajectories using matlab codes we provided in another folder or your own codes.

You can try the optimizer and maybe beat us and publish a new paper.

[arenas7307979]

Hi, @AILocAR

Sorry to bother you again. Recently, I tried training the RouteU training set. I noticed that it contains only two data sets, but the training loss is very high, and the final evaluation differs significantly from the original weights' output result. Could you kindly provide some guidance or advice on this matter??

Thank you!