9c79067ac1768b004ce7a20938d6984a3153f7f5,hypergan/train_hooks/competitive_train_hook.py,CompetitiveTrainHook,gradients,#CompetitiveTrainHook#Any#Any#,108

Before Change

      nsteps = self.config.nsteps or 5
      d_loss, g_loss = self.gan.loss.sample
      d_grads = self.step(0, nsteps, d_grads, d_grads, g_grads, d_loss, g_loss, self.gan.d_vars(), self.gan.g_vars())
      g_grads = self.step(0, nsteps, g_grads, g_grads, d_grads, g_loss, d_loss, self.gan.g_vars(), self.gan.d_vars())
      return [d_grads, g_grads]

  def hvp1(self, grads, x_params, y_params, vs):

After Change

      lr = self.config.learn_rate or 1e-4
      d_grads = self.step(0, nsteps, d_grads, d_grads, g_grads, d_loss, g_loss, d_params, g_params)
      g_grads = self.step(0, nsteps, g_grads, g_grads, d_grads, g_loss, d_loss, g_params, d_params)
      if self.config.final_hvp:
          hvp = self.hvp_function()
          d_grads2 = hvp(g_loss, g_params, d_params, [lr * _g for _g in g_grads])
          g_grads2 = hvp(d_loss, d_params, g_params, [lr * _d for _d in d_grads])
          d_grads = normalize(d_grads2, d_grads, self.config.normalize)
          g_grads = normalize(g_grads2, g_grads, self.config.normalize)
      return [d_grads, g_grads]

  def hvp_function(self):
      hvp = self.hvp

Instances