Regularizer weight in MRKLD

[xiaoaoran]

Regularizer weight in MRKLD

[yzou2]

Thanks for catching up. If I remember correctly, we find the hyperparameters chosen in the repo achieves better performance than the paper. That's why we choose it. The reg_weights are hyperparameters to control if we apply regularization on source or target samples (

CRST/crst_seg.py

Lines 344 to 348 in f7e4df0

	if args.mr_weight_kld == 0:
	reg_weight_tgt = 0.0
	else: # currently only one kind of model regularizer is supported
	reg_weight_tgt = args.mr_weight_kld
	reg_weight_src = args.mr_weight_src

): mr_weight_src = 0 means we do not apply MRKLD regularization in source samples. Same definitions apply to mr_weight_tgt.

Feel free to try to set mr_weight_tgt = 1.0.

[xiaoaoran]

Thank you.