Closed rico227 closed 6 months ago
Thank you for your interest in my work! You should not set the model's type as FrozenBackboneEncoderDecoder. In my code, rein is considered a part of the backbone, thus FrozenBackboneEncoderDecoder would prevent rein from being trained. You should use EncoderDecoder instead. I speculate this will solve your problem and yield reasonable performance. If there are still issues, feel free to continue.
If this work has been helpful to you, please give me a star.😀
Thank you for your interest in my work! You should not set the model's type as FrozenBackboneEncoderDecoder. In my code, rein is considered a part of the backbone, thus FrozenBackboneEncoderDecoder would prevent rein from being trained. You should use EncoderDecoder instead. I speculate this will solve your problem and yield reasonable performance. If there are still issues, feel free to continue.
This makes a lot of sense. Thanks!
If this work has been helpful to you, please give me a star.😀
Already starred! :) Great paper. Very good+interesting read!
I am only getting about 23 mIoU for BDD and 48 for cityscapes, when training on cityscapes using EVA-02-B model + mask2former + rein. Any ideas what could lead to this huge performance drop compared to EVA-L in the paper? Training config is below.