transfer learning

[jiwonpoong]

Dear nnInteractive team,

Thank you very much for developing and sharing nnInteractive.
I've tested it within 3D Slicer and was truly impressed by its powerful segmentation performance.

To conduct more robust and dataset-relevant research, I would like to fine-tune the model on my own data.

Could you please clarify the following:

1. Is it permitted—under the current license—to use the provided pretrained model weights as initialization for fine-tuning?
2. After fine-tuning, can the resulting model weights be used for non-commercial research purposes?
3. If appropriate, would it also be allowed to share the fine-tuned weights with other researchers for academic collaboration?

Thank you for your time and for providing this valuable tool to the community.

[mariamonzon]

Following. Is there also a guide on how to fine-tune the model?

[mrokuss]

Hey @jiwonpoong

Thank you for your kind words and for your interest in nnInteractive. To address your questions:

1. Use of pretrained weights for fine-tuning: Yes, it is permitted to use the provided pretrained weights as initialization for fine-tuning. However, please note that any resulting downstream model must also comply with the non-commercial license.

2. Use of fine-tuned model weights: Yes, the fine-tuned model weights can be used for non-commercial research purposes.

3. Sharing fine-tuned weights for academic collaboration: Yes, sharing the fine-tuned model with other researchers for academic and non-commercial collaboration is allowed under the current license.

Please note that we currently do not provide code for fine-tuning, as it would be essentially identical to the training code. That said, we are actively working on a solution for commercial use and will share updates as soon as they become available.

Best,

Max