print(np.unique(nmi_gts), np.unique(nmi_preds), np.unique(nmi_preds_w_bg))
if is_full:
nmi = normalized_mutual_info_score(nmi_gts, nmi_preds_w_bg) * 100
ari = adjusted_rand_score(nmi_gts, nmi_preds_w_bg) * 100
else:
nmi = normalized_mutual_info_score(nmi_gts, nmi_preds) * 100
ari = adjusted_rand_score(nmi_gts, nmi_preds_w_bg) * 100
return nmi, ari
ari = adjusted_rand_score(nmi_gts, nmi_preds_w_bg) * 100
-->
ari = adjusted_rand_score(nmi_gts, nmi_preds) * 100
in the evaluation notebook the foreground ari score is calculated with nmi predictions with background. FG-ARI on CUB increases from 21.0 to 21.65 when fixed. As this is a main result I think it is important to change this in the code and paper
ari = adjusted_rand_score(nmi_gts, nmi_preds_w_bg) * 100
-->ari = adjusted_rand_score(nmi_gts, nmi_preds) * 100
in the evaluation notebook the foreground ari score is calculated with nmi predictions with background. FG-ARI on CUB increases from 21.0 to 21.65 when fixed. As this is a main result I think it is important to change this in the code and paper