张量流的自定义成本函数

Question

Is there any way I can have this custom cost for my op in Tensorflow?

n_step =5
for i in range(0,int(train_size/n_step)):
    W=[0]
    I=[0]
    for j in range(n_step):
        W.append(max(W[j]+train_y[i*n_step+j]-yhat[i*n_step+j],0))
        I.append(max(-W[j]-train_y[i*n_step+j]+yhat[i*n_step+j],0))

    loss += 2*np.sum(W)+1*np.sum(I) 

where yhat is the output of a fully connected network and W , I are auxiliaries.

Answer 1

Yes, you just have to add your loss to the collection of losses with tf.losses.add_loss .

In order for you to be able to do that, you have to express your loss function in terms of tensors and tensor operations, however. Numpy operations as in your code will not do.