Merge pull request #6 from zombie-einstein/non_static_env_args

Allow non-static env_params and fix changes to reward shape returned …

examples/gym_usage.ipynb

@@ -167,18 +167,19 @@
    "execution_count": null,
    "id": "39e5b9b1",
    "metadata": {
-    "scrolled": true
+    "scrolled": false
    },
    "outputs": [],
    "source": [
-    "_k, trained_agent, losses, ts, rewards = jax_ppo.train(\n",
+    "_k, trained_agent, losses, ts, rewards, _ = jax_ppo.train(\n",
     "    k, env, env_params, agent,\n",
     "    N_TRAIN, \n",
     "    N_TRAIN_ENV, \n",
     "    N_EPOCHS, \n",
     "    MINI_BATCH_SIZE, \n",
     "    N_TEST_ENV, \n",
     "    params, \n",
+    "    env_params.max_steps_in_episode,\n",
     "    greedy_test_policy=True\n",
     ")"
    ]
@@ -190,7 +191,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "plt.plot(jnp.mean(jnp.sum(rewards[:, :, :, 0], axis=2), axis=1));\n",
+    "plt.plot(jnp.mean(jnp.sum(rewards[:, :, :], axis=2), axis=1));\n",
     "plt.xlabel(\"Training Step\")\n",
     "plt.ylabel(\"Avg Total Rewards\");"
    ]
@@ -236,7 +237,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "plt.plot(rewards[-1, :, :, 0].T, drawstyle=\"steps-mid\");\n",
+    "plt.plot(rewards[-1, :, :].T, drawstyle=\"steps-mid\");\n",
     "plt.xlabel(\"Step\");\n",
     "plt.ylabel(\"Reward\");"
    ]

examples/lstm_usage.ipynb

@@ -179,7 +179,7 @@
    },
    "outputs": [],
    "source": [
-    "_k, trained_agent, losses, ts, rewards = jax_ppo.train_recurrent(\n",
+    "_k, trained_agent, losses, ts, rewards, _ = jax_ppo.train_recurrent(\n",
     "    k, env, env_params, agent, \n",
     "    N_TRAIN, \n",
     "    N_TRAIN_ENV, \n",
@@ -190,6 +190,7 @@
     "    1,\n",
     "    N_BURN_IN,\n",
     "    params,\n",
+    "    env_params.max_steps_in_episode,\n",
     "    greedy_test_policy=True,\n",
     ")"
    ]
@@ -201,7 +202,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "plt.plot(jnp.mean(jnp.sum(rewards[:, :, :, 0], axis=2), axis=1));\n",
+    "plt.plot(jnp.mean(jnp.sum(rewards[:, :, :], axis=2), axis=1));\n",
     "plt.xlabel(\"Training Step\")\n",
     "plt.ylabel(\"Avg Total Rewards\");"
    ]
@@ -247,7 +248,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "plt.plot(rewards[-1, :, :, 0].T, drawstyle=\"steps-mid\");\n",
+    "plt.plot(rewards[-1, :, :].T, drawstyle=\"steps-mid\");\n",
     "plt.xlabel(\"Step\");\n",
     "plt.ylabel(\"Reward\");"
    ]

jax_ppo/lstm/training.py

@@ -84,6 +84,7 @@ def _sample_step(carry, _):
         if n_agents is None:
             new_observation = new_observation[jnp.newaxis]
             _done = jnp.array([_done])
+            _reward = jnp.array([_reward])
 
         new_observation = jnp.hstack(
             (_observation.at[:, 1:].get(), new_observation[:, jnp.newaxis])
@@ -96,7 +97,7 @@ def _sample_step(carry, _):
                 action=_action,
                 log_likelihood=_log_likelihood,
                 value=_value,
-                reward=_reward[0],
+                reward=_reward,
                 done=_done,
                 hidden_states=_hidden_state,
             ),
@@ -161,7 +162,7 @@ def _step(carry, _):
 
         return (
             (k, _agent, new_hidden_state, new_state, new_observation),
-            (new_state, _reward[0], _info),
+            (new_state, _reward, _info),
         )
 
     key, observation, state, hidden_states = _reset_env(
@@ -190,7 +191,6 @@ def _step(carry, _):
     jax.jit,
     static_argnames=(
         "env",
-        "env_params",
         "n_train",
         "n_train_env",
         "n_train_epochs",
@@ -218,8 +218,8 @@ def train(
     n_recurrent_layers: int,
     n_burn_in: int,
     ppo_params: data_types.PPOParams,
+    n_env_steps: int,
     n_agents: typing.Optional[int] = None,
-    n_env_steps: typing.Optional[int] = None,
     greedy_test_policy: bool = False,
     max_mini_batches: int = 10_000,
 ) -> typing.Tuple[

jax_ppo/mlp/training.py

@@ -47,6 +47,7 @@ def _sample_step(carry, _):
         if n_agents is None:
             new_observation = new_observation[jnp.newaxis]
             _done = jnp.array([_done])
+            _reward = jnp.array([_reward])
 
         return (
             (k, _agent, new_state, new_observation),
@@ -55,7 +56,7 @@ def _sample_step(carry, _):
                 action=_action,
                 log_likelihood=_log_likelihood,
                 value=_value,
-                reward=_reward[0],
+                reward=_reward,
                 done=_done,
             ),
         )
@@ -118,7 +119,7 @@ def _step(carry, _):
 
         return (
             (k, _agent, new_state, new_observation),
-            (new_state, _reward[0], _info),
+            (new_state, _reward, _info),
         )
 
     key, reset_key = jax.random.split(key)
@@ -138,7 +139,6 @@ def _step(carry, _):
     jax.jit,
     static_argnames=(
         "env",
-        "env_params",
         "n_train",
         "n_train_env",
         "n_train_epochs",
@@ -160,8 +160,8 @@ def train(
     mini_batch_size: int,
     n_test_env: int,
     ppo_params: data_types.PPOParams,
+    n_env_steps: int,
     n_agents: typing.Optional[int] = None,
-    n_env_steps: typing.Optional[int] = None,
     greedy_test_policy: bool = False,
     max_mini_batches: int = 10_000,
 ) -> typing.Tuple[

jax_ppo/runner.py

@@ -26,7 +26,7 @@ def train(
     n_agents: typing.Optional[int],
     greedy_test_policy: bool,
     max_mini_batches: int,
-    n_env_steps: typing.Optional[int],
+    n_env_steps: int,
     **static_kwargs,
 ) -> typing.Tuple[
     jax.random.PRNGKey,
@@ -37,9 +37,6 @@ def train(
     typing.Dict,
 ]:
 
-    if n_env_steps is None:
-        n_env_steps = env_params.max_steps_in_episode
-
     test_keys = jax.random.split(key, n_test_env + 1)
     key, test_keys = test_keys[0], test_keys[1:]
 

tests/conftest.py

@@ -67,7 +67,7 @@ def default_params(self) -> EnvParams:
         def step_env(self, key, state: EnvState, action, params: EnvParams):
             new_state = EnvState(x=state.x + jnp.sum(action) + params.y, t=state.t + 1)
             new_obs = self.get_obs(new_state)
-            rewards = jnp.array([[2.0]])
+            rewards = 2.0
             dones = self.is_terminal(new_state, params)
             return new_obs, new_state, rewards, dones, dict()
 
@@ -124,7 +124,7 @@ def step_env(self, key, state: EnvState, action, params: EnvParams):
                 t=state.t + 1,
             )
             new_obs = self.get_obs(new_state)
-            rewards = jnp.arange(N_AGENTS, dtype=jnp.float32)[jnp.newaxis]
+            rewards = jnp.arange(N_AGENTS, dtype=jnp.float32)
             dones = self.is_terminal(new_state, params)
             return new_obs, new_state, rewards, dones, dict()
 

tests/mlp/test_sampling.py

@@ -63,7 +63,7 @@ def test_shape(x):
         assert x.shape[0] == N_SAMPLES
 
     jax.tree_util.tree_map(test_shape, state_ts)
-    assert reward_ts.shape == (N_SAMPLES, 1)
+    assert reward_ts.shape == (N_SAMPLES,)
 
 
 def test_marl_policy_testing(key, mlp_agent, dummy_marl_env):

tests/recurrent/test_sampling.py

@@ -76,7 +76,7 @@ def test_shape(x):
         assert x.shape[0] == n
 
     jax.tree_util.tree_map(test_shape, state_ts)
-    assert reward_ts.shape == (n, 1)
+    assert reward_ts.shape == (n,)
 
 
 def test_marl_policy_testing(key, recurrent_agent, dummy_marl_env):