🐛 Fixes small learning rate lower bound bug (#105)

* 🔧 Adds submodule pull check to the setup.py

* 🐛 Fixes lr lower bound bug

This commit fixes a small bug which caused a error to be thrown when
trying to fix the learning rate lower bound.

.gitignore

@@ -21,9 +21,8 @@ dist/
 # node related
 node_modules/
 
-# Files to ignore
-.coverage
-*.pyc
+# IDE files to ignore
+.vscode
 *.code-workspace
 
 # Other

machine_learning_control/control/algos/pytorch/lac/lac.py

@@ -411,8 +411,7 @@ def update(self, data):
                 q2_pi_targ = self.ac_targ.Q2(o_, a2)
                 if self._opt_type.lower() == "minimize":
                     q_pi_targ = torch.max(
-                        q1_pi_targ,
-                        q2_pi_targ,
+                        q1_pi_targ, q2_pi_targ,
                     )  # Use max clipping  to prevent overestimation bias.
                 else:
                     q_pi_targ = torch.min(
@@ -656,9 +655,7 @@ def load_state_dict(self, state_dict, restore_lagrance_multipliers=True):
         try:
             super().load_state_dict(state_dict)
         except AttributeError as e:
-            raise type(e)(
-                "The 'state_dict' could not be loaded successfully.",
-            ) from e
+            raise type(e)("The 'state_dict' could not be loaded successfully.",) from e
 
     def _update_targets(self):
         """Updates the target networks based on a Exponential moving average
@@ -721,7 +718,7 @@ def bound_lr(
         if lr_alpha_final is not None:
             if self._log_alpha_optimizer.param_groups[0]["lr"] < lr_a_final:
                 self._log_alpha_optimizer.param_groups[0]["lr"] = lr_a_final
-        if lr_labda_final is not None:
+        if lr_labda_final is not None and self._use_lyapunov:
             if self._log_labda_optimizer.param_groups[0]["lr"] < lr_labda_final:
                 self._log_labda_optimizer.param_groups[0]["lr"] = lr_labda_final
 
@@ -736,8 +733,7 @@ def alpha(self):
     def alpha(self, set_val):
         """Property used to make sure alpha and log_alpha are related."""
         self.log_alpha.data = torch.as_tensor(
-            np.log(1e-37 if set_val < 1e-37 else set_val),
-            dtype=self.log_alpha.dtype,
+            np.log(1e-37 if set_val < 1e-37 else set_val), dtype=self.log_alpha.dtype,
         )
 
     @property
@@ -753,8 +749,7 @@ def labda(self):
     def labda(self, set_val):
         """Property used to make sure labda and log_labda are related."""
         self.log_labda.data = torch.as_tensor(
-            np.log(1e-37 if set_val < 1e-37 else set_val),
-            dtype=self.log_labda.dtype,
+            np.log(1e-37 if set_val < 1e-37 else set_val), dtype=self.log_labda.dtype,
         )
 
     @property
@@ -1179,9 +1174,7 @@ def lac(
                 policy, test_env, num_test_episodes, max_ep_len=max_ep_len
             )
             logger.store(
-                TestEpRet=eps_ret,
-                TestEpLen=eps_len,
-                extend=True,
+                TestEpRet=eps_ret, TestEpLen=eps_len, extend=True,
             )
 
             # Epoch based learning rate decay
@@ -1218,13 +1211,11 @@ def lac(
                 logger.log_tabular("LossQ", average_only=True)
             if adaptive_temperature:
                 logger.log_tabular(
-                    "LossAlpha",
-                    average_only=True,
+                    "LossAlpha", average_only=True,
                 )
             if use_lyapunov:
                 logger.log_tabular(
-                    "LossLambda",
-                    average_only=True,
+                    "LossLambda", average_only=True,
                 )
             if use_lyapunov:
                 logger.log_tabular("LVals", with_min_and_max=True)

machine_learning_control/control/utils/plot.py

@@ -94,7 +94,7 @@ def plot_data(
     Changes the colorscheme and the default legend style, though.
     """
     plt.legend(loc="best").set_draggable(True)
-    # plt.legend(loc='upper center', ncol=3, handlelength=1,
+    # plt.legend(loc='upper center', ncol=6, handlelength=1, mode="expand"
     #           borderaxespad=0., prop={'size': 13})
 
     """

setup.py

@@ -1,12 +1,32 @@
 """Setup file for the 'machine_learning_control' python package.
 """
 
+import os.path as osp
 import re
+import sys
 
 from setuptools import find_namespace_packages, setup
 
+# Script settings
 stand_alone_ns_pkgs = ["simzoo"]
 
+
+def submodules_available(submodules):
+    """Throws warning and stops the script if any of the submodules is not present."""
+    for submodule in submodules:
+        submodule_setup_path = osp.join(
+            osp.abspath(osp.dirname(__file__)),
+            "machine_learning_control",
+            submodule,
+            "setup.py",
+        )
+
+        if not osp.exists(submodule_setup_path):
+            print("Could not find {}".format(submodule_setup_path))
+            print("Did you run 'git submodule update --init --recursive'?")
+            sys.exit(1)
+
+
 # Add extra virtual shortened package for each stand-alone namespace package
 # NOTE: This only works if you don't have a __init__.py file in your parent folder and
 # stand alone_ns_pkgs folder.
@@ -32,9 +52,9 @@
         if short_child not in PACKAGES:
             PACKAGES.append(short_child)
 
+# Throw warning if submodules were not pulled
+submodules_available(stand_alone_ns_pkgs)
+
 setup(
-    packages=PACKAGES,
-    package_dir=PACKAGE_DIR,
+    packages=PACKAGES, package_dir=PACKAGE_DIR,
 )
-
-# TODO: Add submodule check like in pytorch!