Using small training set weights as initialization for large set fine-tuning

[Yuhan Zhou]

Dear Challenge team,

Is it compliant to include weights pre-trained on the small training set in our repository, and have train_model() load and continue fine-tuning on the large training set?

Thank you.

Best,

Yuhan

[PhysioNet Challenge]

Dear Yuhan,

We require that teams submit working training code, and we allow transfer learning as long as it is not an attempt to circumvent this requirement.

If you upload a model that is pre-trained on the small training set and training code that fine-tunes the model on the large training set, then that is fine -- as long as it satisfies this requirement.

For example, if your training code were insensitive to the choice of training data, and it returned approximately the same model whether we trained it on the large training set, a new and different training set, or a completely random training set, then the training code is not working, i.e., it is not learning from the provided training data, which would limit the utility of the approach.

Best,
Matt
(On behalf of the Challenge team.)

Please post questions and comments in the forum. However, if your question reveals information about your entry, then please email info at physionetchallenge.org. We may post parts of our reply publicly if we feel that all Challengers should benefit from it. We will not answer emails about the Challenge to any other address. This email is maintained by a group. Please do not email us individually.