FEATURE: Allows examine_env to load hard coded policies (#87)

* Allow examine_envs to load hard coded policies, E.g.
python robohive/utils/examine_env.py -e FrankaReachRandom-v0 -p robohive.utils.examine_env.rand_policy

* Remove unnecessary import
* examine_env loading scripted polcies: update DESC in examine_env and add unit test
Co-authored-by: Patrick Lancaster <[email protected]>

robohive/tests/test_examine_env.py

@@ -22,10 +22,19 @@ def test_offscreen_rendering(self):
         result = runner.invoke(examine_env, ["--env_name", "door-v1", \
                                             "--num_episodes", 1, \
                                             "--render", "offscreen",\
-                                            "--camera_name", "top_acam"])
+                                            "--camera_name", "top_cam"])
         print(result.output.strip())
         self.assertEqual(result.exception, None)
 
+    def test_scripted_policy_loading(self):
+        # Call your function and test its output/assertions
+        print("Testing scripted policy loading")
+        runner = click.testing.CliRunner()
+        result = runner.invoke(examine_env, ["--env_name", "door-v1", \
+                                            "--num_episodes", 1, \
+                                            "--policy_path", "robohive.utils.examine_env.rand_policy"])
+        print(result.output.strip())
+        self.assertEqual(result.exception, None)
 
 if __name__ == '__main__':
     unittest.main()

robohive/utils/examine_env.py

@@ -16,10 +16,12 @@
 DESC = '''
 Helper script to examine an environment and associated policy for behaviors; \n
 - either onscreen, or offscreen, or just rollout without rendering.\n
-- save resulting paths as pickle or as 2D plots
+- save resulting paths as pickle or as 2D plots \n
+- rollout either learned policies or scripted policies (e.g. see rand_policy class below) \n
 USAGE:\n
-    $ python examine_env.py --env_name door-v0 \n
-    $ python examine_env.py --env_name door-v0 --policy my_policy.pickle --mode evaluation --episodes 10 \n
+    $ python examine_env.py --env_name door-v1 \n
+    $ python examine_env.py --env_name door-v1 --policy_path robohive.utils.examine_env.rand_policy \n
+    $ python examine_env.py --env_name door-v1 --policy_path my_policy.pickle --mode evaluation --episodes 10 \n
 '''
 
 # Random policy
@@ -30,7 +32,14 @@ def __init__(self, env, seed):
 
     def get_action(self, obs):
         # return self.env.np_random.uniform(high=self.env.action_space.high, low=self.env.action_space.low)
-        return self.env.action_space.sample(), {'mode': 'random samples'}
+        return self.env.action_space.sample(), {'mode': 'random samples', 'evaluation':self.env.action_space.sample()}
+
+def load_class_from_str(module_name, class_name):
+    try:
+        m = __import__(module_name, globals(), locals(), class_name)
+        return getattr(m, class_name)
+    except (ImportError, AttributeError):
+        return None
 
 # MAIN =========================================================
 @click.command(help=DESC)
@@ -57,13 +66,19 @@ def main(env_name, policy_path, mode, seed, num_episodes, render, camera_name, o
 
     # resolve policy and outputs
     if policy_path is not None:
-        pi = pickle.load(open(policy_path, 'rb'))
-        if output_dir == './': # overide the default
-            output_dir, pol_name = os.path.split(policy_path)
-            output_name = os.path.splitext(pol_name)[0]
-        if output_name is None:
-            pol_name = os.path.split(policy_path)[1]
-            output_name = os.path.splitext(pol_name)[0]
+        policy_tokens = policy_path.split('.')
+        pi = load_class_from_str('.'.join(policy_tokens[:-1]), policy_tokens[-1])
+
+        if pi is not None:
+            pi = pi(env, seed)
+        else:
+            pi = pickle.load(open(policy_path, 'rb'))
+            if output_dir == './': # overide the default
+                output_dir, pol_name = os.path.split(policy_path)
+                output_name = os.path.splitext(pol_name)[0]
+            if output_name is None:
+                pol_name = os.path.split(policy_path)[1]
+                output_name = os.path.splitext(pol_name)[0]
     else:
         pi = rand_policy(env, seed)
         mode = 'exploration'