def ppo_policy_loss( self, advantages: torch.Tensor, log_probs: torch.Tensor, old_log_probs: torch.Tensor, loss_masks: torch.Tensor, ) -> torch.Tensor: """ Evaluate PPO policy loss. :param advantages: Computed advantages. :param log_probs: Current policy probabilities :param old_log_probs: Past policy probabilities :param loss_masks: Mask for losses. Used with LSTM to ignore 0'ed out experiences. """ advantage = advantages.unsqueeze(-1) decay_epsilon = self.hyperparameters.epsilon r_theta = torch.exp(log_probs - old_log_probs) p_opt_a = r_theta * advantage p_opt_b = ( torch.clamp(r_theta, 1.0 - decay_epsilon, 1.0 + decay_epsilon) * advantage) policy_loss = -1 * ModelUtils.masked_mean(torch.min(p_opt_a, p_opt_b), loss_masks) return policy_loss
def test_tanh_gaussian_dist_instance(): torch.manual_seed(0) act_size = 4 dist_instance = TanhGaussianDistInstance(torch.zeros(1, act_size), torch.ones(1, act_size)) for _ in range(10): action = dist_instance.sample() assert action.shape == (1, act_size) assert torch.max(action) < 1.0 and torch.min(action) > -1.0
def trust_region_policy_loss( advantages: torch.Tensor, log_probs: torch.Tensor, old_log_probs: torch.Tensor, loss_masks: torch.Tensor, epsilon: float, ) -> torch.Tensor: """ Evaluate policy loss clipped to stay within a trust region. Used for PPO and POCA. :param advantages: Computed advantages. :param log_probs: Current policy probabilities :param old_log_probs: Past policy probabilities :param loss_masks: Mask for losses. Used with LSTM to ignore 0'ed out experiences. """ advantage = advantages.unsqueeze(-1) r_theta = torch.exp(log_probs - old_log_probs) p_opt_a = r_theta * advantage p_opt_b = torch.clamp(r_theta, 1.0 - epsilon, 1.0 + epsilon) * advantage policy_loss = -1 * ModelUtils.masked_mean(torch.min(p_opt_a, p_opt_b), loss_masks) return policy_loss
def sac_value_loss( self, log_probs: ActionLogProbs, values: Dict[str, torch.Tensor], q1p_out: Dict[str, torch.Tensor], q2p_out: Dict[str, torch.Tensor], loss_masks: torch.Tensor, ) -> torch.Tensor: min_policy_qs = {} with torch.no_grad(): _cont_ent_coef = self._log_ent_coef.continuous.exp() _disc_ent_coef = self._log_ent_coef.discrete.exp() for name in values.keys(): if self._action_spec.discrete_size <= 0: min_policy_qs[name] = torch.min(q1p_out[name], q2p_out[name]) else: disc_action_probs = log_probs.all_discrete_tensor.exp() _branched_q1p = ModelUtils.break_into_branches( q1p_out[name] * disc_action_probs, self._action_spec.discrete_branches, ) _branched_q2p = ModelUtils.break_into_branches( q2p_out[name] * disc_action_probs, self._action_spec.discrete_branches, ) _q1p_mean = torch.mean( torch.stack([ torch.sum(_br, dim=1, keepdim=True) for _br in _branched_q1p ]), dim=0, ) _q2p_mean = torch.mean( torch.stack([ torch.sum(_br, dim=1, keepdim=True) for _br in _branched_q2p ]), dim=0, ) min_policy_qs[name] = torch.min(_q1p_mean, _q2p_mean) value_losses = [] if self._action_spec.discrete_size <= 0: for name in values.keys(): with torch.no_grad(): v_backup = min_policy_qs[name] - torch.sum( _cont_ent_coef * log_probs.continuous_tensor, dim=1) value_loss = 0.5 * ModelUtils.masked_mean( torch.nn.functional.mse_loss(values[name], v_backup), loss_masks) value_losses.append(value_loss) else: disc_log_probs = log_probs.all_discrete_tensor branched_per_action_ent = ModelUtils.break_into_branches( disc_log_probs * disc_log_probs.exp(), self._action_spec.discrete_branches, ) # We have to do entropy bonus per action branch branched_ent_bonus = torch.stack([ torch.sum(_disc_ent_coef[i] * _lp, dim=1, keepdim=True) for i, _lp in enumerate(branched_per_action_ent) ]) for name in values.keys(): with torch.no_grad(): v_backup = min_policy_qs[name] - torch.mean( branched_ent_bonus, axis=0) # Add continuous entropy bonus to minimum Q if self._action_spec.continuous_size > 0: v_backup += torch.sum( _cont_ent_coef * log_probs.continuous_tensor, dim=1, keepdim=True, ) value_loss = 0.5 * ModelUtils.masked_mean( torch.nn.functional.mse_loss(values[name], v_backup.squeeze()), loss_masks, ) value_losses.append(value_loss) value_loss = torch.mean(torch.stack(value_losses)) if torch.isinf(value_loss).any() or torch.isnan(value_loss).any(): raise UnityTrainerException("Inf found") return value_loss
def sac_value_loss( self, log_probs: torch.Tensor, values: Dict[str, torch.Tensor], q1p_out: Dict[str, torch.Tensor], q2p_out: Dict[str, torch.Tensor], loss_masks: torch.Tensor, discrete: bool, ) -> torch.Tensor: min_policy_qs = {} with torch.no_grad(): _ent_coef = torch.exp(self._log_ent_coef) for name in values.keys(): if not discrete: min_policy_qs[name] = torch.min(q1p_out[name], q2p_out[name]) else: action_probs = log_probs.exp() _branched_q1p = ModelUtils.break_into_branches( q1p_out[name] * action_probs, self.act_size ) _branched_q2p = ModelUtils.break_into_branches( q2p_out[name] * action_probs, self.act_size ) _q1p_mean = torch.mean( torch.stack( [ torch.sum(_br, dim=1, keepdim=True) for _br in _branched_q1p ] ), dim=0, ) _q2p_mean = torch.mean( torch.stack( [ torch.sum(_br, dim=1, keepdim=True) for _br in _branched_q2p ] ), dim=0, ) min_policy_qs[name] = torch.min(_q1p_mean, _q2p_mean) value_losses = [] if not discrete: for name in values.keys(): with torch.no_grad(): v_backup = min_policy_qs[name] - torch.sum( _ent_coef * log_probs, dim=1 ) value_loss = 0.5 * ModelUtils.masked_mean( torch.nn.functional.mse_loss(values[name], v_backup), loss_masks ) value_losses.append(value_loss) else: branched_per_action_ent = ModelUtils.break_into_branches( log_probs * log_probs.exp(), self.act_size ) # We have to do entropy bonus per action branch branched_ent_bonus = torch.stack( [ torch.sum(_ent_coef[i] * _lp, dim=1, keepdim=True) for i, _lp in enumerate(branched_per_action_ent) ] ) for name in values.keys(): with torch.no_grad(): v_backup = min_policy_qs[name] - torch.mean( branched_ent_bonus, axis=0 ) value_loss = 0.5 * ModelUtils.masked_mean( torch.nn.functional.mse_loss(values[name], v_backup.squeeze()), loss_masks, ) value_losses.append(value_loss) value_loss = torch.mean(torch.stack(value_losses)) if torch.isinf(value_loss).any() or torch.isnan(value_loss).any(): raise UnityTrainerException("Inf found") return value_loss