chain.py 文件源码

def _update_d_and_f(self, state):
        d, f = state['d'], state['f']
        if self.t > 1:
            old_f = float(cuda.to_cpu(state['f']))
            if self.loss > old_f:
                delta = self.lower_threshold + 1.
                Delta = self.upper_threshold + 1.
            else:
                delta = 1. / (self.upper_threshold + 1.)
                Delta = 1. / (self.lower_threshold + 1.)
            c = min(max(delta, self.loss / (old_f + 1e-12)), Delta)
            new_f = c * old_f
            r = abs(new_f - old_f) / (min(new_f, old_f) + 1e-12)
            d += (1 - self.beta3) * (r - d)
            f[:] = new_f
        else:
            f[:] = self.loss