[Feature] flexible batch_locked for jumanji

ghstack-source-id: 61e86da575685838059aac81e216731f80b40832
Pull Request resolved: #2382

test/test_libs.py

@@ -1553,7 +1553,7 @@ def test_jumanji_seeding(self, envname):
 
     @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
     def test_jumanji_batch_size(self, envname, batch_size):
-        env = JumanjiEnv(envname, batch_size=batch_size)
+        env = JumanjiEnv(envname, batch_size=batch_size, jit=True)
         env.set_seed(0)
         tdreset = env.reset()
         tdrollout = env.rollout(max_steps=50)
@@ -1564,7 +1564,7 @@ def test_jumanji_batch_size(self, envname, batch_size):
 
     @pytest.mark.parametrize("batch_size", [(), (5,), (5, 4)])
     def test_jumanji_spec_rollout(self, envname, batch_size):
-        env = JumanjiEnv(envname, batch_size=batch_size)
+        env = JumanjiEnv(envname, batch_size=batch_size, jit=True)
         env.set_seed(0)
         check_env_specs(env)
 
@@ -1575,7 +1575,7 @@ def test_jumanji_consistency(self, envname, batch_size):
         import numpy as onp
         from torchrl.envs.libs.jax_utils import _tree_flatten
 
-        env = JumanjiEnv(envname, batch_size=batch_size)
+        env = JumanjiEnv(envname, batch_size=batch_size, jit=True)
         obs_keys = list(env.observation_spec.keys(True))
         env.set_seed(1)
         rollout = env.rollout(10)
@@ -1613,7 +1613,7 @@ def test_jumanji_consistency(self, envname, batch_size):
     @pytest.mark.parametrize("batch_size", [[3], []])
     def test_jumanji_rendering(self, envname, batch_size):
         # check that this works with a batch-size
-        env = JumanjiEnv(envname, from_pixels=True, batch_size=batch_size)
+        env = JumanjiEnv(envname, from_pixels=True, batch_size=batch_size, jit=True)
         env.set_seed(0)
         env.transform.transform_observation_spec(env.base_env.observation_spec)
 
@@ -1626,6 +1626,23 @@ def test_jumanji_rendering(self, envname, batch_size):
 
         check_env_specs(env)
 
+    @pytest.mark.parametrize("jit", [True, False])
+    def test_jumanji_batch_unlocked(self, envname, jit):
+        torch.manual_seed(0)
+        env = JumanjiEnv(envname, jit=jit)
+        env.set_seed(0)
+        assert not env.batch_locked
+        reset = env.reset(TensorDict(batch_size=[16]))
+        assert reset.batch_size == (16,)
+        env.rand_step(reset)
+        t0 = time.time()
+        r = env.rollout(
+            20, auto_reset=False, tensordict=reset, break_when_all_done=True
+        )
+        assert r.batch_size[0] == 16
+        done = r["next", "done"].float()
+        assert (done.cumprod(-2) == done).all()
+
 
 ENVPOOL_CLASSIC_CONTROL_ENVS = [
     PENDULUM_VERSIONED(),

torchrl/envs/common.py

@@ -1473,6 +1473,16 @@ def step(self, tensordict: TensorDictBase) -> TensorDictBase:
         """
         # sanity check
         self._assert_tensordict_shape(tensordict)
+        if not self.batch_locked:
+            # Batched envs have their own way of dealing with this - batched envs that are not batched-locked may fail here
+            partial_steps = tensordict.get("_step", None)
+            if partial_steps is not None:
+                if partial_steps.all():
+                    partial_steps = None
+                else:
+                    tensordict_batch_size = tensordict.batch_size
+                    partial_steps = partial_steps.view(tensordict_batch_size)
+                    tensordict = tensordict[partial_steps]
         next_preset = tensordict.get("next", None)
 
         next_tensordict = self._step(tensordict)
@@ -1485,6 +1495,10 @@ def step(self, tensordict: TensorDictBase) -> TensorDictBase:
                 next_preset.exclude(*next_tensordict.keys(True, True))
             )
         tensordict.set("next", next_tensordict)
+        if partial_steps is not None:
+            result = tensordict.new_zeros(tensordict_batch_size)
+            result[partial_steps] = tensordict
+            return result
         return tensordict
 
     @classmethod
@@ -2696,7 +2710,7 @@ def _rollout_stop_early(
             if break_when_all_done:
                 if partial_steps is not True:
                     # At least one partial step has been done
-                    del td_append["_partial_steps"]
+                    del td_append["_step"]
                     td_append = torch.where(
                         partial_steps.view(td_append.shape), td_append, tensordicts[-1]
                     )
@@ -2722,17 +2736,17 @@ def _rollout_stop_early(
                 _terminated_or_truncated(
                     tensordict,
                     full_done_spec=self.output_spec["full_done_spec"],
-                    key="_partial_steps",
+                    key="_step",
                     write_full_false=False,
                 )
-                partial_step_curr = tensordict.get("_partial_steps", None)
+                partial_step_curr = tensordict.get("_step", None)
                 if partial_step_curr is not None:
                     partial_step_curr = ~partial_step_curr
                     partial_steps = partial_steps & partial_step_curr
                 if partial_steps is not True:
                     if not partial_steps.any():
                         break
-                    tensordict.set("_partial_steps", partial_steps)
+                    tensordict.set("_step", partial_steps)
 
             if callback is not None:
                 callback(self, tensordict)

torchrl/envs/libs/jax_utils.py

@@ -102,19 +102,21 @@ def _object_to_tensordict(obj, device, batch_size) -> TensorDictBase:
     return None
 
 
-def _tensordict_to_object(tensordict: TensorDictBase, object_example):
+def _tensordict_to_object(tensordict: TensorDictBase, object_example, batch_size=None):
     """Converts a TensorDict to a namedtuple or a dataclass."""
     from jax import dlpack as jax_dlpack, numpy as jnp
 
+    if batch_size is None:
+        batch_size = []
     t = {}
     _fields = _get_object_fields(object_example)
     for name, example in _fields.items():
         value = tensordict.get(name, None)
         if isinstance(value, TensorDictBase):
-            t[name] = _tensordict_to_object(value, example)
+            t[name] = _tensordict_to_object(value, example, batch_size=batch_size)
         elif value is None:
             if isinstance(example, dict):
-                t[name] = _tensordict_to_object({}, example)
+                t[name] = _tensordict_to_object({}, example, batch_size=batch_size)
             else:
                 t[name] = None
         else:
@@ -140,7 +142,9 @@ def _tensordict_to_object(tensordict: TensorDictBase, object_example):
                     t[name] = value
             else:
                 value = jnp.reshape(value, tuple(shape))
-                t[name] = value.view(example.dtype).reshape(example.shape)
+                t[name] = value.view(example.dtype).reshape(
+                    (*batch_size, *example.shape)
+                )
     return type(object_example)(**t)
 
 

torchrl/envs/libs/jumanji.py

@@ -128,6 +128,12 @@ class JumanjiWrapper(GymLikeEnv, metaclass=_JumanjiMakeRender):
 
     Paper: https://arxiv.org/abs/2306.09884
 
+    .. note:: For better performance, turn `jit` on when instantiating this class.
+        The `jit` attribute can also be flipped during code execution:
+
+            >>> env.jit = True # Used jit
+            >>> env.jit = False # eager
+
     Args:
         env (jumanji.env.Environment): the env to wrap.
         categorical_action_encoding (bool, optional): if ``True``, categorical
@@ -136,6 +142,22 @@ class JumanjiWrapper(GymLikeEnv, metaclass=_JumanjiMakeRender):
             Defaults to ``False``.
 
     Keyword Args:
+        batch_size (torch.Size, optional): the batch size of the environment.
+            With ``jumanji``, this indicates the number of vectorized environments.
+            If the batch-size is empty, the environment is not batch-locked and an arbitrary number
+            of environments can be executed simultaneously.
+            Defaults to ``torch.Size([])``.
+
+                >>> import jumanji
+                >>> from torchrl.envs import JumanjiWrapper
+                >>> base_env = jumanji.make("Snake-v1")
+                >>> env = JumanjiWrapper(base_env)
+                >>> # Set the batch-size of the TensorDict instead of the env allows to control the number
+                >>> #  of envs being run simultaneously
+                >>> tdreset = env.reset(TensorDict(batch_size=[32]))
+                >>> # Execute a rollout until all envs are done or max steps is reached, whichever comes first
+                >>> rollout = env.rollout(100, break_when_all_done=True, auto_reset=False, tensordict=tdreset)
+
         from_pixels (bool, optional): Whether the environment should render its output.
             This will drastically impact the environment throughput. Only the first environment
             will be rendered. See :meth:`~torchrl.envs.JumanjiWrapper.render` for more information.
@@ -146,17 +168,15 @@ class JumanjiWrapper(GymLikeEnv, metaclass=_JumanjiMakeRender):
             of rewards across steps.
         device (torch.device, optional): if provided, the device on which the data
             is to be cast. Defaults to ``torch.device("cpu")``.
-        batch_size (torch.Size, optional): the batch size of the environment.
-            With ``jumanji``, this indicates the number of vectorized environments.
-            Defaults to ``torch.Size([])``.
         allow_done_after_reset (bool, optional): if ``True``, it is tolerated
             for envs to be ``done`` just after :meth:`~.reset` is called.
             Defaults to ``False``.
+        jit (bool, optional): whether the step and reset method should be wrapped in `jit`.
+            Defaults to ``False``.
 
     Attributes:
         available_envs: environments availalbe to build
 
-    Examples:
     Examples:
         >>> import jumanji
         >>> from torchrl.envs import JumanjiWrapper
@@ -334,6 +354,7 @@ def __init__(
         self,
         env: "jumanji.env.Environment" = None,  # noqa: F821
         categorical_action_encoding=True,
+        jit: bool = True,
         **kwargs,
     ):
         if not _has_jumanji:
@@ -343,7 +364,26 @@ def __init__(
         self.categorical_action_encoding = categorical_action_encoding
         if env is not None:
             kwargs["env"] = env
+        batch_locked = kwargs.pop("batch_locked", kwargs.get("batch_size") is not None)
         super().__init__(**kwargs)
+        self._batch_locked = batch_locked
+        self.jit = jit
+
+    @property
+    def jit(self):
+        return self._jit
+
+    @jit.setter
+    def jit(self, value):
+        self._jit = value
+        if value:
+            import jax
+
+            self._env_reset = jax.jit(self._env.reset)
+            self._env_step = jax.jit(self._env.step)
+        else:
+            self._env_reset = self._env.reset
+            self._env_step = self._env.step
 
     def _build_env(
         self,
@@ -486,17 +526,21 @@ def _set_seed(self, seed):
             raise Exception("Jumanji requires an integer seed.")
         self.key = jax.random.PRNGKey(seed)
 
-    def read_state(self, state):
-        state_dict = _object_to_tensordict(state, self.device, self.batch_size)
+    def read_state(self, state, batch_size=None):
+        state_dict = _object_to_tensordict(
+            state, self.device, self.batch_size if batch_size is None else batch_size
+        )
         return self.state_spec["state"].encode(state_dict)
 
-    def read_obs(self, obs):
+    def read_obs(self, obs, batch_size=None):
         from jax import numpy as jnp
 
         if isinstance(obs, (list, jnp.ndarray, np.ndarray)):
             obs_dict = _ndarray_to_tensor(obs).to(self.device)
         else:
-            obs_dict = _object_to_tensordict(obs, self.device, self.batch_size)
+            obs_dict = _object_to_tensordict(
+                obs, self.device, self.batch_size if batch_size is None else batch_size
+            )
         return super().read_obs(obs_dict)
 
     def render(
@@ -561,7 +605,11 @@ def render(
             isinteractive = plt.isinteractive()
             plt.ion()
             buf = io.BytesIO()
-            state = _tensordict_to_object(tensordict.get("state"), _state_example)
+            state = _tensordict_to_object(
+                tensordict.get("state"),
+                _state_example,
+                batch_size=tensordict.batch_size if not self.batch_locked else None,
+            )
             self._env.render(state, **kwargs)
             plt.savefig(buf, format="png")
             buf.seek(0)
@@ -580,24 +628,33 @@ def render(
     def _step(self, tensordict: TensorDictBase) -> TensorDictBase:
         import jax
 
+        if self.batch_locked:
+            batch_size = self.batch_size
+        else:
+            batch_size = tensordict.batch_size
+
         # prepare inputs
-        state = _tensordict_to_object(tensordict.get("state"), self._state_example)
+        state = _tensordict_to_object(
+            tensordict.get("state"),
+            self._state_example,
+            batch_size=tensordict.batch_size if not self.batch_locked else None,
+        )
         action = self.read_action(tensordict.get("action"))
 
         # flatten batch size into vector
-        state = _tree_flatten(state, self.batch_size)
-        action = _tree_flatten(action, self.batch_size)
+        state = _tree_flatten(state, batch_size)
+        action = _tree_flatten(action, batch_size)
 
         # jax vectorizing map on env.step
-        state, timestep = jax.vmap(self._env.step)(state, action)
+        state, timestep = jax.vmap(self._env_step)(state, action)
 
         # reshape batch size from vector
-        state = _tree_reshape(state, self.batch_size)
-        timestep = _tree_reshape(timestep, self.batch_size)
+        state = _tree_reshape(state, batch_size)
+        timestep = _tree_reshape(timestep, batch_size)
 
         # collect outputs
-        state_dict = self.read_state(state)
-        obs_dict = self.read_obs(timestep.observation)
+        state_dict = self.read_state(state, batch_size=batch_size)
+        obs_dict = self.read_obs(timestep.observation, batch_size=batch_size)
         reward = self.read_reward(np.asarray(timestep.reward))
         done = timestep.step_type == self.lib.types.StepType.LAST
         done = _ndarray_to_tensor(done).view(torch.bool).to(self.device)
@@ -622,32 +679,63 @@ def _reset(
         import jax
         from jax import numpy as jnp
 
+        if self.batch_locked:
+            numel = self.numel()
+            batch_size = self.batch_size
+        else:
+            numel = tensordict.numel()
+            batch_size = tensordict.batch_size
+
         # generate random keys
-        self.key, *keys = jax.random.split(self.key, self.numel() + 1)
+        self.key, *keys = jax.random.split(self.key, numel + 1)
 
         # jax vectorizing map on env.reset
-        state, timestep = jax.vmap(self._env.reset)(jnp.stack(keys))
+        state, timestep = jax.vmap(self._env_reset)(jnp.stack(keys))
 
         # reshape batch size from vector
-        state = _tree_reshape(state, self.batch_size)
-        timestep = _tree_reshape(timestep, self.batch_size)
+        state = _tree_reshape(state, batch_size)
+        timestep = _tree_reshape(timestep, batch_size)
 
         # collect outputs
-        state_dict = self.read_state(state)
-        obs_dict = self.read_obs(timestep.observation)
-        done_td = self.full_done_spec.zero()
+        state_dict = self.read_state(state, batch_size=batch_size)
+        obs_dict = self.read_obs(timestep.observation, batch_size=batch_size)
+        if not self.batch_locked:
+            done_td = self.full_done_spec.zero(batch_size)
+        else:
+            done_td = self.full_done_spec.zero()
 
         # build results
         tensordict_out = TensorDict(
             source=obs_dict,
-            batch_size=self.batch_size,
+            batch_size=batch_size,
             device=self.device,
         )
         tensordict_out.update(done_td)
         tensordict_out["state"] = state_dict
 
         return tensordict_out
 
+    def read_reward(self, reward):
+        """Reads the reward and maps it to the reward space.
+
+        Args:
+            reward (torch.Tensor or TensorDict): reward to be mapped.
+
+        """
+        if isinstance(reward, int) and reward == 0:
+            return self.reward_spec.zero()
+        if self.batch_locked:
+            reward = self.reward_spec.encode(reward, ignore_device=True)
+        else:
+            reward = torch.as_tensor(reward)
+            if reward.shape[-1] != self.reward_spec.shape[-1]:
+                reward = reward.unsqueeze(-1)
+
+        if reward is None:
+            reward = torch.tensor(np.nan).expand(self.reward_spec.shape)
+
+        return reward
+
     def _output_transform(self, step_outputs_tuple: Tuple) -> Tuple:
         ...