MambaBDA-Small performs better than MambaBDA-Base after finetuning on xBD

[rrryan2016]

The best result of MambaBDA-Small:

---------starting validation-----------
Validation: lofF1 is 0.0, oaF1 is 0.5303396099452545, final_OA:0.8840646534920277, clfF1(harmonic_mean_f1) is 0.7576280142075065, sub class F1 score is [0.95039659 0.59241224 0.74248799 0.83849895], mmd_f1:0.7244663931021744,IoU_of_each_class:[0. 0.9054817 0.4208706 0.59044213 0.72190984], mIoU:0.5277408551235119, harmonic_mean_iou:4.999996720366557e-07, IoU_mmd:0.5777408573160148
---------starting test-----------
Test: lofF1 is 0.0, oaF1 is 0.5251892418982564, final_OA:0.8888830094723688, clfF1(harmonic_mean_f1) is 0.7502703455689379, sub class F1 score is [0.94936371 0.57652762 0.73802356 0.84133821], mmd_f1:0.7186297972576705,IoU_of_each_class:[0. 0.90360843 0.40501501 0.58481571 0.72612932], mIoU:0.5239136936140252, harmonic_mean_iou:4.999996668590888e-07, IoU_mmd:0.5719866807179388

The best result of MambaBDA-Base:

iter is 90000, localization loss is 0.0, classification loss is 1.2516288757324219
---------starting validation-----------
Validation: lofF1 is 0.0, oaF1 is 0.5259345456142492, final_OA:0.8742176043625283, clfF1(harmonic_mean_f1) is 0.7513350651632132, sub class F1 score is [0.94530994 0.58362301 0.73823461 0.8347348 ], mmd_f1:0.7188641414507161,IoU_of_each_class:[0. 0.89629179 0.41205349 0.58508079 0.71634756], mIoU:0.5219547269095701, harmonic_mean_iou:4.999996676146035e-07, IoU_mmd:0.5711606149920697
---------starting test-----------
Test: lofF1 is 0.0, oaF1 is 0.5172245519555511, final_OA:0.879073173902418, clfF1(harmonic_mean_f1) is 0.7388922170793587, sub class F1 score is [0.9443129 0.55789898 0.72326411 0.84783183], mmd_f1:0.7096649730031288,IoU_of_each_class:[0. 0.89450084 0.38686544 0.56649476 0.73585779], mIoU:0.516743764657174, harmonic_mean_iou:4.999996586493141e-07, IoU_mmd:0.5630726611342166

Does it make sense?

Btw, would you like share the checkpoint on xBD?