Function EasyTL

code/traditional/pyEasyTL/EasyTL.py:74–121

(Xs,Ys,Xt,Yt,intra_align="coral",dist="euclidean",lp="linear")

    return source_class_center, Dct

def EasyTL(Xs,Ys,Xt,Yt,intra_align="coral",dist="euclidean",lp="linear"):
# Inputs:
#   Xs          : source data, ns * m
#   Ys          : source label, ns * 1
#   Xt          : target data, nt * m
#   Yt          : target label, nt * 1
# The following inputs are not necessary
#   intra_align : intra-domain alignment: coral(default)|gfk|pca|raw
#   dist        : distance: Euclidean(default)|ma(Mahalanobis)|cosine|rbf
#   lp          : linear(default)|binary

# Outputs:
#   acc         : final accuracy
#   y_pred      : predictions for target domain
    
# Reference:
# Jindong Wang, Yiqiang Chen, Han Yu, Meiyu Huang, Qiang Yang.
# Easy Transfer Learning By Exploiting Intra-domain Structures.
# IEEE International Conference on Multimedia & Expo (ICME) 2019.

    C = len(np.unique(Ys))
    if C > np.max(Ys):
        Ys += 1
        Yt += 1
	
    m = len(Yt)
	
    if intra_align == "raw":
        print('EasyTL using raw feature...')
    elif intra_align == "pca":
        print('EasyTL using PCA...')
        print('Not implemented yet, using raw feature')
		#Xs, Xt = PCA_map(Xs, Xt)
    elif intra_align == "gfk":
        print('EasyTL using GFK...')
        print('Not implemented yet, using raw feature')
        #Xs, Xt = GFK_map(Xs, Xt)
    elif intra_align == "coral":
        print('EasyTL using CORAL...')
        Xs = CORAL_map(Xs, Xt)
	
    _, Dct = get_class_center(Xs,Ys,Xt,dist)
    print('Start intra-domain programming...')
    Mcj = label_prop(C,m,Dct,lp)
    y_pred = np.argmax(Mcj, axis=1) + 1
    acc = np.mean(y_pred == Yt.flatten());

    return acc, y_pred