Sync predict with SB and add version file

docs/misc/changelog.rst

@@ -11,6 +11,7 @@ Breaking Changes:
 ^^^^^^^^^^^^^^^^^
 - Removed default seed
 - Bump dependencies (PyTorch and Gym)
+- ``predict()`` now returns a tuple to match Stable-Baselines behavior
 
 New Features:
 ^^^^^^^^^^^^^
@@ -32,6 +33,7 @@ Others:
 - More typing
 - Add test for ``expln``
 - Renamed ``learning_rate`` to ``lr_schedule``
+- Add ``version.txt``
 
 Documentation:
 ^^^^^^^^^^^^^^

setup.py

@@ -1,7 +1,11 @@
+import os
 import sys
 import subprocess
 from setuptools import setup, find_packages
 
+with open(os.path.join('torchy_baselines', 'version.txt'), 'r') as file_handler:
+    __version__ = file_handler.read()
+
 
 setup(name='torchy_baselines',
       packages=[package for package in find_packages()
@@ -48,7 +52,7 @@ setup(name='torchy_baselines',
       license="MIT",
       long_description="",
       long_description_content_type='text/markdown',
-      version="0.2.4",
+      version=__version__,
       )
 
 # python setup.py sdist

tests/test_predict.py

@@ -34,10 +34,10 @@ def test_predict(model_class, env_id):
     vec_env = DummyVecEnv([lambda: gym.make(env_id), lambda: gym.make(env_id)])
 
     obs = env.reset()
-    action = model.predict(obs)
+    action, _ = model.predict(obs)
     assert action.shape == env.action_space.shape
     assert env.action_space.contains(action)
 
     vec_env_obs = vec_env.reset()
-    action = model.predict(vec_env_obs)
+    action, _ = model.predict(vec_env_obs)
     assert action.shape[0] == vec_env_obs.shape[0]

tests/test_save_load.py

@@ -54,7 +54,7 @@ def test_save_load(model_class):
     params = new_params
 
     # get selected actions
-    selected_actions = model.predict(observations, deterministic=True)
+    selected_actions, _ = model.predict(observations, deterministic=True)
 
     # Check
     model.save("test_save.zip")
@@ -69,7 +69,7 @@ def test_save_load(model_class):
         assert th.allclose(params[key], new_params[key]), "Model parameters not the same after save and load."
 
     # check if model still selects the same actions
-    new_selected_actions = model.predict(observations, deterministic=True)
+    new_selected_actions, _ = model.predict(observations, deterministic=True)
     assert np.allclose(selected_actions, new_selected_actions, 1e-4)
 
     # check if learn still works

torchy_baselines/__init__.py

@@ -1,7 +1,12 @@
+import os
+
 from torchy_baselines.a2c import A2C
 from torchy_baselines.cem_rl import CEMRL
 from torchy_baselines.ppo import PPO
 from torchy_baselines.sac import SAC
 from torchy_baselines.td3 import TD3
 
-__version__ = "0.2.4"
+# Read version from file
+version_file = os.path.join(os.path.dirname(__file__), 'version.txt')
+with open(version_file, 'r') as file_handler:
+    __version__ = file_handler.read()

torchy_baselines/common/base_class.py

@@ -369,7 +369,7 @@ class BaseRLModel(ABC):
     def predict(self, observation: np.ndarray,
                 state: Optional[np.ndarray] = None,
                 mask: Optional[np.ndarray] = None,
-                deterministic: bool = False) -> np.ndarray:
+                deterministic: bool = False) -> Tuple[np.ndarray, Optional[np.ndarray]]:
         """
         Get the model's action(s) from an observation
 
@@ -377,7 +377,7 @@ class BaseRLModel(ABC):
         :param state: (Optional[np.ndarray]) The last states (can be None, used in recurrent policies)
         :param mask: (Optional[np.ndarray]) The last masks (can be None, used in recurrent policies)
         :param deterministic: (bool) Whether or not to return deterministic actions.
-        :return: (np.ndarray) the model's action and the next state (used in recurrent policies)
+        :return: (Tuple[np.ndarray, Optional[np.ndarray]]) the model's action and the next state (used in recurrent policies)
         """
         # if state is None:
         #     state = self.initial_state
@@ -409,9 +409,7 @@ class BaseRLModel(ABC):
                 raise ValueError("Error: The environment must be vectorized when using recurrent policies.")
             clipped_actions = clipped_actions[0]
 
-        # TODO: switch to stable baselines API
-        # return clipped_actions, state
-        return clipped_actions
+        return clipped_actions, state
 
     @classmethod
     def load(cls, load_path: str, env: Optional[GymEnv] = None, **kwargs):
@@ -896,7 +894,7 @@ class OffPolicyRLModel(BaseRLModel):
                 else:
                     # Note: we assume that the policy uses tanh to scale the action
                     # We use non-deterministic action in the case of SAC, for TD3, it does not matter
-                    unscaled_action = self.predict(obs, deterministic=False)
+                    unscaled_action, _ = self.predict(obs, deterministic=False)
 
                 # Rescale the action from [low, high] to [-1, 1]
                 scaled_action = self.scale_action(unscaled_action)

torchy_baselines/common/evaluation.py

@@ -8,11 +8,11 @@ def evaluate_policy(model, env, n_eval_episodes=10, deterministic=True,
                     render=False, callback=None, reward_threshold=None,
                     return_episode_rewards=False):
     """
-    Runs policy for `n_eval_episodes` episodes and returns average reward.
+    Runs policy for ``n_eval_episodes`` episodes and returns average reward.
     This is made to work only with one env.
 
     :param model: (BaseRLModel) The RL agent you want to evaluate.
-    :param env: (gym.Env or VecEnv) The gym environment. In the case of a `VecEnv`
+    :param env: (gym.Env or VecEnv) The gym environment. In the case of a ``VecEnv``
         this must contain only one environment.
     :param n_eval_episodes: (int) Number of episode to evaluate the agent
     :param deterministic: (bool) Whether to use deterministic or stochastic actions
@@ -24,7 +24,7 @@ def evaluate_policy(model, env, n_eval_episodes=10, deterministic=True,
     :param return_episode_rewards: (bool) If True, a list of reward per episode
         will be returned instead of the mean.
     :return: (float, float) Mean reward per episode, std of reward per episode
-        returns ([float], [int]) when `return_episode_rewards` is True
+        returns ([float], [int]) when ``return_episode_rewards`` is True
     """
     if isinstance(env, VecEnv):
         assert env.num_envs == 1, "You must pass only one environment when using this function"
@@ -32,11 +32,11 @@ def evaluate_policy(model, env, n_eval_episodes=10, deterministic=True,
     episode_rewards, episode_lengths = [], []
     for _ in range(n_eval_episodes):
         obs = env.reset()
-        done = False
+        done, state = False, None
         episode_reward = 0.0
         episode_length = 0
         while not done:
-            action = model.predict(obs, deterministic=deterministic)
+            action, state = model.predict(obs, state=state, deterministic=deterministic)
             obs, reward, done, _info = env.step(action)
             episode_reward += reward
             if callback is not None:

torchy_baselines/version.txt

@@ -0,0 +1 @@
+0.2.5