447444fd06594e531ae1141afac78051481e4468,catalyst/rl/offpolicy/algorithms/td3.py,TD3,_base_loss,#TD3#Any#Any#Any#Any#Any#,107

Before Change

        policy_loss = -torch.mean(q_values_tp0_min)

        // critic loss
        actions_tp1 = self.target_actor(states_tp1).detach()
        actions_tp1 = self._add_noise_to_actions(actions_tp1)
        q_values_t = [
            x(states_t, actions_t).squeeze_(dim=3)
            for x in self.critics

After Change

        // [bs; num_heads; 1] -> many-heads view transform
        // [{bs * num_heads}; 1]
        q_target_t = (
            rewards_t + (1 - done_t) * gammas * q_values_tp1
        ).view(-1, 1).detach()

        value_loss = [
            self.critic_criterion(x, q_target_t).mean() for x in q_values_t
        ]

Instances