Clean-up: Removed unused variables and fixed typo

sbx/common/jax_layers.py

@@ -162,8 +162,6 @@ def __call__(self, x, use_running_average: Optional[bool] = None):
             custom_mean = mean
             custom_var = var
             if not self.is_initializing():
-                r = jnp.array(1.0)
-                d = jnp.array(0.0)
                 std = jnp.sqrt(var + self.epsilon)
                 ra_std = jnp.sqrt(ra_var.value + self.epsilon)
                 # scale

sbx/crossq/crossq.py

@@ -208,7 +208,6 @@ def train(self, gradient_steps: int, batch_size: int) -> None:
             (actor_loss_value, qf_loss_value, ent_coef_value),
         ) = self._train(
             self.gamma,
-            self.tau,
             self.target_entropy,
             gradient_steps,
             data,
@@ -260,7 +259,7 @@ def mse_loss(
             #
             # This has two reasons:
             # 1. According to the paper obs/actions and next_obs/next_state_actions are differently
-            #    distributed which is the reason why "naively" appling Batch Normalization in SAC fails.
+            #    distributed which is the reason why "naively" applying Batch Normalization in SAC fails.
             #    The batch statistics have to instead be calculated for the mixture distribution of obs/next_obs
             #    and actions/next_state_actions. Otherwise, next_obs/next_state_actions are perceived as
             #    out-of-distribution to the Batch Normalization layer, since running statistics are only polyak averaged
@@ -385,7 +384,6 @@ def update_actor_and_temperature(
     def _train(
         cls,
         gamma: float,
-        tau: float,
         target_entropy: ArrayLike,
         gradient_steps: int,
         data: ReplayBufferSamplesNp,