Update SB3 ONNX export documentation

docs/guide/export.rst

@@ -31,53 +31,52 @@ to do inference in another framework.
 Export to ONNX
 -----------------
 
-As of June 2021, ONNX format  `doesn't support <https://github.com/onnx/onnx/issues/3033>`_ exporting models that use the ``broadcast_tensors`` functionality of pytorch. So in order to export the trained stable-baseline3 models in the ONNX format, we need to first remove the layers that use broadcasting. This can be done by creating a class that removes the unsupported layers.
 
-The following examples are for ``MlpPolicy`` only, and are general examples. Note that you have to preprocess the observation the same way stable-baselines3 agent does (see ``common.preprocessing.preprocess_obs``).
+If you are using PyTorch 2.0+ and ONNX Opset 14+, you can easily export SB3 policies using the following code:
 
-For PPO, assuming a shared feature extractor.
 
 .. warning::
 
-  The following example is for continuous actions only.
-  When using discrete or binary actions, you must do some `post-processing <https://github.com/DLR-RM/stable-baselines3/blob/f3a35aa786ee41ffff599b99fa1607c067e89074/stable_baselines3/common/policies.py#L621-L637>`_
-  to obtain the action (e.g., convert action logits to action).
+  The following returns normalized actions and doesn't include the `post-processing <https://github.com/DLR-RM/stable-baselines3/blob/a9273f968eaf8c6e04302a07d803eebfca6e7e86/stable_baselines3/common/policies.py#L370-L377>`_ step that is done with continuous actions
+  (clip or unscale the action to the correct space).
 
 
 .. code-block:: python
 
   import torch as th
+  from typing import Tuple
 
   from stable_baselines3 import PPO
+  from stable_baselines3.common.policies import BasePolicy
 
 
-  class OnnxablePolicy(th.nn.Module):
-      def __init__(self, extractor, action_net, value_net):
+  class OnnxableSB3Policy(th.nn.Module):
+      def __init__(self, policy: BasePolicy):
           super().__init__()
-          self.extractor = extractor
-          self.action_net = action_net
-          self.value_net = value_net
+          self.policy = policy
 
-      def forward(self, observation):
-          # NOTE: You may have to process (normalize) observation in the correct
-          #       way before using this. See `common.preprocessing.preprocess_obs`
-          action_hidden, value_hidden = self.extractor(observation)
-          return self.action_net(action_hidden), self.value_net(value_hidden)
+      def forward(self, observation: th.Tensor) -> Tuple[th.Tensor, th.Tensor, th.Tensor]:
+          # NOTE: Preprocessing is included, but postprocessing
+          # (clipping/inscaling actions) is not,
+          # If needed, you also need to transpose the images so that they are channel first
+          # use deterministic=False if you want to export the stochastic policy
+          # policy() returns `actions, values, log_prob` for PPO
+          return self.policy(observation, deterministic=True)
 
 
   # Example: model = PPO("MlpPolicy", "Pendulum-v1")
+  PPO("MlpPolicy", "Pendulum-v1").save("PathToTrainedModel")
   model = PPO.load("PathToTrainedModel.zip", device="cpu")
-  onnxable_model = OnnxablePolicy(
-      model.policy.mlp_extractor, model.policy.action_net, model.policy.value_net
-  )
+
+  onnx_policy = OnnxableSB3Policy(model.policy)
 
   observation_size = model.observation_space.shape
   dummy_input = th.randn(1, *observation_size)
   th.onnx.export(
-      onnxable_model,
+      onnx_policy,
       dummy_input,
       "my_ppo_model.onnx",
-      opset_version=9,
+      opset_version=17,
       input_names=["input"],
   )
 
@@ -93,7 +92,13 @@ For PPO, assuming a shared feature extractor.
 
   observation = np.zeros((1, *observation_size)).astype(np.float32)
   ort_sess = ort.InferenceSession(onnx_path)
-  action, value = ort_sess.run(None, {"input": observation})
+  actions, values, log_prob = ort_sess.run(None, {"input": observation})
+
+  print(actions, values, log_prob)
+
+  # Check that the predictions are the same
+  with th.no_grad():
+      print(model.policy(th.as_tensor(observation), deterministic=True))
 
 
 For SAC the procedure is similar. The example shown only exports the actor network as the actor is sufficient to roll out the trained policies.
@@ -108,23 +113,16 @@ For SAC the procedure is similar. The example shown only exports the actor netwo
   class OnnxablePolicy(th.nn.Module):
       def __init__(self, actor: th.nn.Module):
           super().__init__()
-          # Removing the flatten layer because it can't be onnxed
-          self.actor = th.nn.Sequential(
-              actor.latent_pi,
-              actor.mu,
-              # For gSDE
-              # th.nn.Hardtanh(min_val=-actor.clip_mean, max_val=actor.clip_mean),
-              # Squash the output
-              th.nn.Tanh(),
-          )
+          self.actor = actor
 
       def forward(self, observation: th.Tensor) -> th.Tensor:
-          # NOTE: You may have to process (normalize) observation in the correct
-          #       way before using this. See `common.preprocessing.preprocess_obs`
-          return self.actor(observation)
+          # NOTE: You may have to postprocess (unnormalize) actions
+          # to the correct bounds (see commented code below)
+          return self.actor(observation, deterministic=True)
 
 
   # Example: model = SAC("MlpPolicy", "Pendulum-v1")
+  SAC("MlpPolicy", "Pendulum-v1").save("PathToTrainedModel.zip")
   model = SAC.load("PathToTrainedModel.zip", device="cpu")
   onnxable_model = OnnxablePolicy(model.policy.actor)
 
@@ -134,7 +132,7 @@ For SAC the procedure is similar. The example shown only exports the actor netwo
       onnxable_model,
       dummy_input,
       "my_sac_actor.onnx",
-      opset_version=9,
+      opset_version=17,
       input_names=["input"],
   )
 
@@ -147,10 +145,23 @@ For SAC the procedure is similar. The example shown only exports the actor netwo
 
   observation = np.zeros((1, *observation_size)).astype(np.float32)
   ort_sess = ort.InferenceSession(onnx_path)
-  action = ort_sess.run(None, {"input": observation})
+  scaled_action = ort_sess.run(None, {"input": observation})[0]
+
+  print(scaled_action)
+
+  # Post-process: rescale to correct space
+  # Rescale the action from [-1, 1] to [low, high]
+  # low, high = model.action_space.low, model.action_space.high
+  # post_processed_action = low + (0.5 * (scaled_action + 1.0) * (high - low))
+
+  # Check that the predictions are the same
+  with th.no_grad():
+      print(model.actor(th.as_tensor(observation), deterministic=True))
+
+
+For more discussion around the topic, please refer to `GH#383 <https://github.com/DLR-RM/stable-baselines3/issues/383>`_ and `GH#1349 <https://github.com/DLR-RM/stable-baselines3/issues/1349>`_.
 
 
-For more discussion around the topic refer to this `issue. <https://github.com/DLR-RM/stable-baselines3/issues/383>`_
 
 Trace/Export to C++
 -------------------

docs/guide/rl_tips.rst

@@ -252,6 +252,12 @@ A better solution would be to use a squashing function (cf ``SAC``) or a Beta di
 Tips and Tricks when implementing an RL algorithm
 =================================================
 
+.. note::
+
+  We have a `video on YouTube about reliable RL <https://www.youtube.com/watch?v=7-PUg9EAa3Y>`_ that covers
+  this section in more details. You can also find the `slides online <https://araffin.github.io/slides/tips-reliable-rl/>`_.
+
+
 When you try to reproduce a RL paper by implementing the algorithm, the `nuts and bolts of RL research <http://joschu.net/docs/nuts-and-bolts.pdf>`_
 by John Schulman are quite useful (`video <https://www.youtube.com/watch?v=8EcdaCk9KaQ>`_).
 
@@ -282,4 +288,4 @@ in RL with discrete actions:
 3. Pong (one of the easiest Atari game)
 4. other Atari games (e.g. Breakout)
 
-.. _SBX: https://github.com/araffin/sbx
+.. _SBX: https://github.com/araffin/sbx

docs/misc/changelog.rst

@@ -59,6 +59,8 @@ Documentation:
 ^^^^^^^^^^^^^^
 - Added a paragraph on modifying vectorized environment parameters via setters (@fracapuano)
 - Updated callback code example
+- Updated export to ONNX documentation, it is now much simpler to export SB3 models with newer ONNX Opset!
+- Added video link to "Practical Tips for Reliable Reinforcement Learning" video
 
 Release 2.2.1 (2023-11-17)
 --------------------------