[Announcement] Regarding Transfer Learning!


Some participants have inquired about the possibility of doing transfer learning for the task. After an internal team meeting, we have decided to allow transfer learning in terms of pre-training. It means that you can use any data to pre-train your models and use those pre-trained weights to initialize your submitted models. More details on the limitation of the submitted checkpoints (initialization weights) will follow [@mohanty].


Hello, can I please ask: can we submit checkpoints, or do you mean we must initialize our weights by submitting data to conduct pre-training on your server?


@jriseley : You can include pre-trained weights in your submission, you will just have to ensure to use git-lfs to check in the model weights.

and the second part of the question is more complicated. As long as you use standard pre-trained models to get started in your training phase, I think its still fair.