fixed spelling

cleanrl/ppo_trxl/ppo_trxl.py

@@ -183,7 +183,7 @@ def forward(self, values, keys, query, mask):
         if mask is not None:
             energy = energy.masked_fill(mask.unsqueeze(1).unsqueeze(1) == 0, float("-1e20"))  # -inf causes NaN
 
-        # Normalize energy values and apply softmax to retreive the attention scores
+        # Normalize energy values and apply softmax to retrieve the attention scores
         attention = torch.softmax(
             energy / (self.embed_dim ** (1 / 2)), dim=3
         )  # attention shape: (N, heads, query_len, key_len)
@@ -387,7 +387,7 @@ def reconstruct_observation(self):
         max_episode_steps = envs.envs[0].max_episode_steps
     if max_episode_steps <= 0:
         max_episode_steps = 1024  # Memory Gym envs have max_episode_steps set to -1
-    # Set transformer memory length to max episode steps if greather than max episode steps
+    # Set transformer memory length to max episode steps if greater than max episode steps
     args.trxl_memory_length = min(args.trxl_memory_length, max_episode_steps)
 
     agent = Agent(args, observation_space, action_space_shape, max_episode_steps).to(device)

docs/rl-algorithms/ppo-trxl.md

@@ -74,7 +74,7 @@ Most details are derived from [ppo.py](/rl-algorithms/ppo#ppopy). These are addi
 
 1. The policy and value function share parameters.
 2. Multi-head attention is implemented so that all heads share parameters.
-3. Abolute positional encoding is used as default. Learned positional encodings are supported.
+3. Absolute positional encoding is used as default. Learned positional encodings are supported.
 4. Previously computed hidden states of the TrXL layers are cached and re-used for up to `trxl_memory_length`. Only 1 hidden state is computed anew.
 5. TrXL layers adhere to pre-layer normalization.
 6. Support for multi-discrete action spaces.