beautiful loss function

最初模型，收益增加９％

def mycrossentropy(y_true, y_pred, e=0.001):
    print('y_pred',y_pred)
    print('y_true',y_true)
    b=y_pred[:,:n_classes]
    b1=y_pred[:,n_classes:]
    print('b=',b)
    print('b1',b1)    
    c=y_true[:,:n_classes]
    c1=y_true[:,n_classes:]
    print('c=',c)
    print('c1',c1)

    loss1 = K.categorical_crossentropy(b, c)
    loss2 =K.abs( K.sum(b1) -K.sum(c1) )

    return   loss1*(loss2**3/6+loss2**2/2+loss2+1)

第二模型，收益增加＊％

def mycrossentropy(y_true, y_pred, e=0.001):
    print('y_pred',y_pred)
    print('y_true',y_true)
    b=y_pred[:,:n_classes]
    b1=y_pred[:,n_classes:]
    print('b=',b)
    print('b1',b1)
##        
    c=y_true[:,:n_classes]
    c1=y_true[:,n_classes:]
    print('c=',c)
    print('c1',c1)
#    b=Dense(n_classes, activation='softmax',name="b")(b)
#
    loss1 = K.categorical_crossentropy(b, c)

#        loss2 = K.(K.ones_like(y_pred)/nb_classes, y_pred)
#    loss2 =  K.sum(K.square(b1 -c1), axis=-1)
    l2 =  K.sum(K.pow((b1 -c1),2), axis=-1)
    l3 =  K.sum(K.pow((b1 -c1),4), axis=-1)
    l4 =  K.sum(K.pow((b1 -c1),6), axis=-1)
    
    
#    loss2 =K.abs( K.sum(b1) -K.sum(c1) )
#    loss2 =K.square( K.sum(b1) -K.sum(c1) )
#        loss2 =  K.categorical_crossentropy(b, y_true)
#    return   loss1
#    return   loss1*(loss2**3.6+loss2**2+loss2)
    return   loss1*(l2/2+l3/6+l4/24)

def mycrossentropy(y_true, y_pred, e=0.001):
    print('y_pred',y_pred)
    print('y_true',y_true)
    b=y_pred[:,:n_classes]
    b1=y_pred[:,n_classes:]
    print('b=',b)
    print('b1',b1)
##        
    c=y_true[:,:n_classes]
    c1=y_true[:,n_classes:]
    print('c=',c)
    print('c1',c1)

    loss1 = K.categorical_crossentropy(b, c)


    ab1 =  K.sum(K.pow(K.abs(b1-c1),1), axis=-1)
    ab2 =  K.sum(K.pow(K.abs(b1-c1),2), axis=-1)
    ab3 =  K.sum(K.pow(K.abs(b1-c1),3), axis=-1)
    ab4 =  K.sum(K.pow(K.abs(b1-c1),4), axis=-1)
    ab5 =  K.sum(K.pow(K.abs(b1-c1),5), axis=-1)

    
  
    return   loss1*(1+ab1+ab2/2+ab3/6+ab4/24+ab5/120)