Using pre-trained models

[Mohammad Hossein Behboudi]

Hi,

Can we use the pre-trained models that already exist (not trained by us), and include their weights in the repo?

I found this in FAQ:
Am I allowed to do transfer learning using pre-trained networks?

Yes, most certainly. We encourage you to do this. You do not need to include your data in the code stack for training the algorithm, but you do need to include the pre-trained model in the code and provide code to retrain (continue training) on the training data we provide. You must also thoroughly document the content of the database you used to pre-train. You cannot use transfer learning to avoid training your model.

But in this forum, you said:
In particular, we do not allow teams to upload pre-processed data or a pre-trained model to avoid these constraints, which support the development of reproducible approaches, allow us to provide more timely feedback, and provide a more even playing field for teams

[PhysioNet Challenge]

Dear Mohammad,

Yes, you can include pre-trained models in your submission for transfer learning, but not to avoid the training requirements and constraints.

For example, a team can include a pre-trained model from a different training set for a different task to perform transfer learning on the Challenge data, but they cannot include a pre-trained model on the Challenge data to avoid providing working training code that we can run on the Challenge data within the training time limit. If the model produced by the team's training code changes minimally or not at all with when the provided training set changes by a non-minimal amount, then the training code is not working.

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.