BP4D_EAC_dlib_train.py 文件源码

def multi_label_ACE(outputs,y_labels):
    data_shape=outputs.shape
    loss_buff=0
    # num=T.iscalar(data_shape[0]) #theano int to get value from tensor
    # for i in range(int(num)):
    #     for j in range(12):
    #         y_exp=outputs[i,j]
    #         y_tru=y_labels[i,0,0,j]
    #         if y_tru==0:
    #             loss_ij=math.log(1-outputs[i,j])
    #             loss_buff-=loss_ij
    #         if y_tru>0:
    #             loss_ij=math.log(outputs[i,j])
    #             loss_buff-=loss_ij

    # wts=[ 0.24331649,  0.18382575,  0.23082499,  0.44545567,  0.52901483,  0.58482504, \
    # 0.57321465,  0.43411294,  0.15502839,  0.36377019,  0.19050646,  0.16083916]
    # for i in [3,4,5,6,7,9]:

    for i in range(12):
        target=y_labels[:,i]
        output=outputs[:,i]
        loss_au=T.sum(-(target * T.log((output+0.05)/1.05) + (1.0 - target) * T.log((1.05 - output)/1.05)))
        loss_buff+=loss_au
    return loss_buff/(12*BATCH_SIZE)