search

fullflu / learning-to-rank

learning to rank repository
26 stars 4 forks source link

loss function compute problem #4

Open speeding-motor opened 5 years ago

speeding-motor commented 5 years ago

HI, when I read you code , it make me refused here, in the the LIstNet loss function:

def get_loss(self, x_t, y_t):
        # ---- start loss calculation ----
        ...
        p_true = F.softmax(F.reshape(y_t,(y_t.shape[0],y_t.shape[1])))
        xm = F.max(pred,axis=1,keepdims = True)
        logsumexp = F.logsumexp(pred,axis=1)

        logsumexp = F.broadcast_to(logsumexp,(xm.shape[0],pred.shape[1]))
        loss = -1 * F.sum( p_true * (pred - logsumexp) )
        ...

here use p_true * (red- logsumexp), why do like this, you are exactly not follow the paper to compute the loss function right?

speeding-motor commented 5 years ago

HI, when I read you code , it make me refused here, in the the LIstNet loss function:

def get_loss(self, x_t, y_t):
        # ---- start loss calculation ----
        ...
        p_true = F.softmax(F.reshape(y_t,(y_t.shape[0],y_t.shape[1])))
        xm = F.max(pred,axis=1,keepdims = True)
        logsumexp = F.logsumexp(pred,axis=1)

        logsumexp = F.broadcast_to(logsumexp,(xm.shape[0],pred.shape[1]))
        loss = -1 * F.sum( p_true * (pred - logsumexp) )
        ...

here use p_true * (red- logsumexp), why do like this, you are exactly not follow the paper to compute the loss function right?

But why do you calculate the loss function in this way? is that work well?