Mismatch pretrained model and checkpoint about resvmamba HOT 3 CLOSED

@DianaZhongHua , thanks for reading our work! You did't misunderstand anything, the model class is still 1000 is because we only change the data file to the CNFOOD-241, the data file folder structure is the same as the imagenet, but only input 241 classes of data.
If you has some reproduced result, it is welcome to share with us!

from resvmamba.

Hi ChiSheng,
Thank you for your prompt response to my previous query about using the CNFOOD-241 dataset with the 1000-class model structure. I'm curious about the specifics of how the 241 classes are mapped to the model's 1000-class output layer.

Could you shed some light on how you handle the mapping during training and inference? Are there certain indices within the 1000 classes that are designated for CNFOOD-241, or do you apply any post-processing to filter out the predictions for the non-existing classes?

Any details you can provide would be greatly helpful for my understanding. Thank you!

from resvmamba.

Hi @DianaZhongHua , the reason that the class input in model is 1000 is because I use the original imagenet dataloader but input the CNFOOD241 data folder, and according to my understanding, although the input class is 1000, but the model weight still can train with the classes under the 1000 (like 241) still works fine. You can see the build.py for detail:

I have re-run the code, and it worked fine in my environment on a 3080:

from resvmamba.