suppress errors in reagent

Differential Revision: D47142615

fbshipit-source-id: 5a6e3a5e3fc202dd5cdc890d1508a57d58b6cef4

Author: generatedunixname89002005307016

reagent/evaluation/evaluation_data_page.py

@@ -317,24 +317,17 @@ def create_from_tensors_dqn(
         metrics: Optional[torch.Tensor] = None,
     ):
         old_q_train_state = trainer.q_network.training
-        # pyre-fixme[16]: `DQNTrainer` has no attribute `reward_network`.
         old_reward_train_state = trainer.reward_network.training
-        # pyre-fixme[16]: Item `Tensor` of `Union[Tensor, Module]` has no attribute
-        #  `training`.
         old_q_cpe_train_state = trainer.q_network_cpe.training
         trainer.q_network.train(False)
-        # pyre-fixme[16]: `Tensor` has no attribute `train`.
         trainer.reward_network.train(False)
-        # pyre-fixme[16]: Item `Tensor` of `Union[Tensor, Module]` has no attribute
-        #  `train`.
         trainer.q_network_cpe.train(False)
 
         num_actions = trainer.num_actions
         action_mask = actions.float()
 
         # pyre-fixme[6]: Expected `Tensor` for 2nd param but got `FeatureData`.
         rewards = trainer.boost_rewards(rewards, actions)
-        # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function.
         model_values = trainer.q_network_cpe(states)[:, 0:num_actions]
         # TODO: make generic get_action_idxs for each trainer class
         # Note: model_outputs are obtained from the q_network for DQN algorithms
@@ -360,7 +353,6 @@ def create_from_tensors_dqn(
             + str(possible_actions_mask.shape)
         )
 
-        # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function.
         rewards_and_metric_rewards = trainer.reward_network(states)
 
         # In case we reuse the modular for Q-network
@@ -390,7 +382,6 @@ def create_from_tensors_dqn(
             model_metrics_for_logged_action = None
             model_metrics_values_for_logged_action = None
         else:
-            # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function.
             model_metrics_values = trainer.q_network_cpe(states)
             # Backward compatility
             if hasattr(model_metrics_values, "q_values"):
@@ -430,12 +421,8 @@ def create_from_tensors_dqn(
                 model_metrics_values_for_logged_action_list, dim=1
             )
 
-        # pyre-fixme[16]: Item `Tensor` of `Union[Tensor, Module]` has no attribute
-        #  `train`.
         trainer.q_network_cpe.train(old_q_cpe_train_state)
         trainer.q_network.train(old_q_train_state)
-        # pyre-fixme[16]: Item `Tensor` of `Union[Tensor, Module]` has no attribute
-        #  `train`.
         trainer.reward_network.train(old_reward_train_state)
 
         return cls(

reagent/gym/policies/scorers/continuous_scorer.py

@@ -11,7 +11,6 @@ def sac_scorer(actor_network: ModelBase) -> Scorer:
     @torch.no_grad()
     def score(preprocessed_obs: rlt.FeatureData) -> GaussianSamplerScore:
         actor_network.eval()
-        # pyre-fixme[29]: `Union[torch.Tensor, torch.nn.Module]` is not a function.
         loc, scale_log = actor_network._get_loc_and_scale_log(preprocessed_obs)
         actor_network.train()
         return GaussianSamplerScore(loc=loc, scale_log=scale_log)

reagent/model_managers/discrete_dqn_base.py

@@ -86,8 +86,6 @@ def create_policy(
             )
         else:
             sampler = GreedyActionSampler()
-            # pyre-fixme[6]: Expected `ModelBase` for 1st param but got
-            #  `Union[torch.Tensor, torch.nn.Module]`.
             scorer = discrete_dqn_scorer(trainer_module.q_network)
             return Policy(scorer=scorer, sampler=sampler)
 

reagent/model_managers/parametric_dqn_base.py

@@ -72,7 +72,6 @@ def create_policy(
         """Create an online DiscreteDQN Policy from env."""
 
         # FIXME: this only works for one-hot encoded actions
-        # pyre-fixme[16]: `Tensor` has no attribute `input_prototype`.
         action_dim = trainer_module.q_network.input_prototype()[1].float_features.shape[
             1
         ]
@@ -87,8 +86,6 @@ def create_policy(
             sampler = SoftmaxActionSampler(temperature=self.rl_parameters.temperature)
             scorer = parametric_dqn_scorer(
                 max_num_actions=action_dim,
-                # pyre-fixme[6]: Expected `ModelBase` for 2nd param but got
-                #  `Union[torch.Tensor, torch.nn.Module]`.
                 q_network=trainer_module.q_network,
             )
             return Policy(scorer=scorer, sampler=sampler)

reagent/model_managers/slate_q_base.py

@@ -72,8 +72,6 @@ def create_policy(
         else:
             scorer = slate_q_scorer(
                 num_candidates=self.num_candidates,
-                # pyre-fixme[6]: Expected `ModelBase` for 2nd param but got
-                #  `Union[torch.Tensor, torch.nn.Module]`.
                 q_network=trainer_module.q_network,
             )
             sampler = TopKSampler(k=self.slate_size)

reagent/models/seq2slate.py

@@ -856,7 +856,6 @@ class Seq2SlateNet(ModelBase):
 
     def __post_init_post_parse__(self) -> None:
         super().__init__()
-        # pyre-fixme[16]: `Seq2SlateNet` has no attribute `seq2slate`.
         self.seq2slate = self._build_model()
 
     def _build_model(self):
@@ -879,7 +878,6 @@ def forward(
         greedy: Optional[bool] = None,
     ):
         if mode == Seq2SlateMode.RANK_MODE:
-            # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function.
             res = self.seq2slate(
                 mode=mode.value,
                 state=input.state.float_features,
@@ -899,7 +897,6 @@ def forward(
             assert input.tgt_in_seq is not None
             assert input.tgt_in_idx is not None
             assert input.tgt_out_idx is not None
-            # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function.
             res = self.seq2slate(
                 mode=mode.value,
                 state=input.state.float_features,
@@ -915,7 +912,6 @@ def forward(
             return rlt.RankingOutput(log_probs=log_probs)
         elif mode == Seq2SlateMode.ENCODER_SCORE_MODE:
             assert input.tgt_out_idx is not None
-            # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function.
             res = self.seq2slate(
                 mode=mode.value,
                 state=input.state.float_features,

reagent/net_builder/synthetic_reward_net_builder.py

@@ -68,8 +68,6 @@ def build_serving_module(
                 seq_len,
                 state_preprocessor,
                 action_preprocessor,
-                # pyre-fixme[29]: `Union[torch.Tensor, torch.nn.Module]` is not a
-                #  function.
                 synthetic_reward_network.export_mlp().cpu().eval(),
             )
         else:

reagent/training/cfeval/bayes_by_backprop_trainer.py

@@ -16,7 +16,6 @@ def train_step_gen(
     ):
         weight = self._get_sample_weight(training_batch)
 
-        # pyre-ignore seems to be pyre bug for pytorch
         loss = self.reward_net.sample_elbo(
             torch.cat([training_batch.action, training_batch.state.float_features], 1),
             training_batch.reward,
@@ -48,7 +47,6 @@ def validation_step(self, batch: rlt.BanditRewardModelInput, batch_idx: int):
             batch = self._training_batch_type.from_dict(batch)
 
         weight = self._get_sample_weight(batch)
-        # pyre-ignore
         loss = self.reward_net.sample_elbo(
             torch.cat([batch.action, batch.state.float_features], 1),
             batch.reward,

reagent/training/dqn_trainer.py

@@ -231,7 +231,6 @@ def compute_td_loss(
 
         # Get Q-value of action taken
         all_q_values = self.q_network(batch.state)
-        # pyre-fixme[16]: `DQNTrainer` has no attribute `all_action_scores`.
         self.all_action_scores = all_q_values.detach()
         q_values = torch.sum(all_q_values * batch.action, 1, keepdim=True)
         td_loss = self.q_network_loss(q_values, target_q_values.detach())

reagent/training/dqn_trainer_base.py

@@ -266,22 +266,17 @@ def _initialize_cpe(
             optimizer: an optimizer object for training q_network_cpe.
         """
         if not self.calc_cpe_in_training:
-            # pyre-fixme[16]: `DQNTrainerBase` has no attribute `reward_network`.
             self.reward_network = None
             return
 
         assert reward_network is not None, "reward_network is required for CPE"
         self.reward_network = reward_network
-        # pyre-fixme[16]: `DQNTrainerBase` has no attribute `reward_network_optimizer`.
         self.reward_network_optimizer = optimizer
         assert (
             q_network_cpe is not None and q_network_cpe_target is not None
         ), "q_network_cpe and q_network_cpe_target are required for CPE"
-        # pyre-fixme[16]: `DQNTrainerBase` has no attribute `q_network_cpe`.
         self.q_network_cpe = q_network_cpe
-        # pyre-fixme[16]: `DQNTrainerBase` has no attribute `q_network_cpe_target`.
         self.q_network_cpe_target = q_network_cpe_target
-        # pyre-fixme[16]: `DQNTrainerBase` has no attribute `q_network_cpe_optimizer`.
         self.q_network_cpe_optimizer = optimizer
         num_output_nodes = len(self.metrics_to_score) * self.num_actions
         reward_idx_offsets = torch.arange(
@@ -295,7 +290,6 @@ def _initialize_cpe(
         reward_stripped_metrics_to_score = (
             self.metrics_to_score[:-1] if len(self.metrics_to_score) > 1 else None
         )
-        # pyre-fixme[16]: `DQNTrainerBase` has no attribute `evaluator`.
         self.evaluator = Evaluator(
             self._actions,
             self.rl_parameters.gamma,