f2ede1d376601f4f6c905d57ef240e1457f1d00b,chainerrl/agents/c51.py,C51,_compute_loss,#C51#Any#Any#Any#,66

Before Change

        

        self.n_atoms = 51
        self.v_min = 0//-10
        self.v_max = 500/ቦ
        self.delta_z = (self.v_max - self.v_min) / float(self.n_atoms - 1)
        self.z_values = self.xp.array([self.v_min + i * self.delta_z for i in range(self.n_atoms)])

        y, t = self._compute_y_and_t(exp_batch, gamma)

        loss = -F.mean(F.sum(t * F.log(y+1e-5), axis=1))

After Change

    def _compute_loss(self, exp_batch, gamma, errors_out=None):
        Compute a loss of categorical DQN.
        y, t = self._compute_y_and_t(exp_batch, gamma)
        eltwise_loss = -t * F.log(y)
        if self.batch_accumulator == "sum":
            loss = F.sum(eltwise_loss)
        else:
            loss = F.mean(F.sum(eltwise_loss, axis=1))
        return loss

Instances