I recently started using SAM model on my baseline, what I have done can be said to be a custom implementation although I do use the encoder weights from the officially trained SAM github where applicable.
If I am finetuning my predictions using SAM, does every model of mine that makes the final prediction (the prediction which goes into the submission .zip file) have to SAM?, or can I ensemble non-SAM and SAM model predictions?
Hi @harshitsheoranlearnt. This is an interesting idea. Can you please send a more detailed description of your use of a SAM/non-SAM ensemble model to challenges@thinkonward.com. Our team will need to review to fully understand how SAM will be used. We will try to respond quickly since tomorrow is the challenge deadline.
Thanks,
Onward Team
Thank you for considering it, I have sent a write-up summarizing my solution.
Update: I just got 0.62 by ensembling all types of models