Thanks for your excellent work. I met some troubles when I try to write scripts to use the "m_t.py", which is the implementation of the Eq.(3) in your article.
(1) What data structure should the cfg.DIC_FILE (the confounder set C) and the cfg.PRIOR_PROB (the P(c)) be ? As what I understand, the cfg.DIC_FILE is a .npy file which contains a hxwx21 numpy array, and the value of each channel (hxw) is in the range of 0~1. And the cfg.PRIOR_PROB is a .npy file which contains a 1x1x21 numpy array, and every element is set to 1/n. Could you please tell me if I understand this correctly?
(2) What the parameter "proposals" represent in the "m_t.py" and how to obtain it ?
(3) The return of "m_t.py" is a causal_logits_list, which seems to be used for calculate loss, so where should this CausalPredictor be added and what loss should be used (add a new loss or use the original multilabel_soft_margin_loss) for the next round classification?
Thanks for your excellent work. I met some troubles when I try to write scripts to use the "m_t.py", which is the implementation of the Eq.(3) in your article.
(1) What data structure should the cfg.DIC_FILE (the confounder set C) and the cfg.PRIOR_PROB (the P(c)) be ? As what I understand, the cfg.DIC_FILE is a .npy file which contains a hxwx21 numpy array, and the value of each channel (hxw) is in the range of 0~1. And the cfg.PRIOR_PROB is a .npy file which contains a 1x1x21 numpy array, and every element is set to 1/n. Could you please tell me if I understand this correctly?
(2) What the parameter "proposals" represent in the "m_t.py" and how to obtain it ?
(3) The return of "m_t.py" is a causal_logits_list, which seems to be used for calculate loss, so where should this CausalPredictor be added and what loss should be used (add a new loss or use the original multilabel_soft_margin_loss) for the next round classification?
Thanks!