fix(torch): handle 'update_after' set to zero (#408)

This commit addresses a bug in the learning rate decay logic when 'update_after' is set to zero. Previously, the algorithm would malfunction under these conditions. With this fix, the algorithm can now correctly handle 'update_after' being set to zero, ensuring proper learning rate decay.
rickstaa · Feb 12, 2024 · 7999590 · 7999590
1 parent 642a193
commit 7999590
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 0 deletions.
diff --git a/stable_learning_control/algos/pytorch/lac/lac.py b/stable_learning_control/algos/pytorch/lac/lac.py
@@ -1021,6 +1021,7 @@ def lac(
             -   replay_buffer (union[:class:`~stable_learning_control.algos.pytorch.common.buffers.ReplayBuffer`, :class:`~stable_learning_control.algos.pytorch.common.buffers.FiniteHorizonReplayBuffer`]):
                 The replay buffer used during training.
     """  # noqa: E501, D301
+    update_after = max(1, update_after)  # You can not update before the first step.
     validate_args(**locals())
 
     # Retrieve hyperparameters while filtering out the logger_kwargs.

diff --git a/stable_learning_control/algos/pytorch/sac/sac.py b/stable_learning_control/algos/pytorch/sac/sac.py
@@ -889,6 +889,7 @@ def sac(
             -   replay_buffer (union[:class:`~stable_learning_control.algos.common.buffers.ReplayBuffer`, :class:`~stable_learning_control.algos.common.buffers.FiniteHorizonReplayBuffer`]):
                 The replay buffer used during training.
     """  # noqa: E501, D301
+    update_after = max(1, update_after)  # You can not update before the first step.
     validate_args(**locals())
 
     # Retrieve hyperparameters while filtering out the logger_kwargs.