from ray.rllib.algorithms.ppo import PPOConfig
config = PPOConfig()

Even though the new API stack still provides rudimentary support for `TensorFlow <https://tensorflow.org>`__,
RLlib supports a single deep learning framework, the `PyTorch <https://pytorch.org>`__
framework, dropping TensorFlow support entirely.
Note, though, that the Ray team continues to design RLlib to be framework-agnostic
and may add support for additional frameworks in the future.

.. testcode::

    config.api_stack(
        enable_rl_module_and_learner=False,
        enable_env_runner_and_connector_v2=False,
    )

Even though the new API stack still provides rudimentary support for `TensorFlow <https://tensorflow.org>`__,
RLlib supports a single deep learning framework, the `PyTorch <https://pytorch.org>`__ framework.

The new API stack deprecates the following framework-related settings:

.. testcode::

    # Make sure you always set the framework to "torch"...
    config.framework("torch")

    # ... and drop all tf-specific settings.
    config.framework(
        eager_tracing=True,
        eager_max_retraces=20,
        tf_session_args={},
        local_tf_session_args={},
    )


AlgorithmConfig.resources()

# The following setting is equivalent to the old stack's `config.resources(num_gpus=2)`.
config.learners(
    num_learners=2,
    num_gpus_per_learner=1,
)

The `num_learners` setting determines how many remote :py:class:`~ray.rllib.core.learner.learner.Learner`
workers there are in your Algorithm's :py:class:`~ray.rllib.core.learner.learner_group.LearnerGroup`.
If you set this parameter to ``0``, your LearnerGroup only contains a **local** Learner that runs on the main
process and shares its compute resources, typically 1 CPU.
For asynchronous algorithms like IMPALA or APPO, this setting should therefore always be >0.

config.learners(
    num_learners=2,  # or >2
    num_cpus_per_learner=1,  # <- default
    num_gpus_per_learner=0,  # <- default
)

config.resources(num_cpus_for_main_process=0)  # default is 1

Train batch size
................

Due to the new API stack's :py:class:`~ray.rllib.core.learner.learner.Learner` worker architecture,
training may happen in distributed fashion over ``n`` :py:class:`~ray.rllib.core.learner.learner.Learner` workers,
so RLlib provides the train batch size per individual :py:class:`~ray.rllib.core.learner.learner.Learner`.
Don't use the ``train_batch_size`` setting any longer:


.. testcode::

    config.training(
        train_batch_size_per_learner=512,
    )

You don't need to change this setting, even when increasing the number of
:py:class:`~ray.rllib.core.learner.learner.Learner`, through `config.learners(num_learners=...)`.

Note that a good rule of thumb for scaling on the learner axis is to keep the
`train_batch_size_per_learner` value constant with a growing number of Learners and
to increase the learning rate as follows:

`lr = [original_lr] * ([num_learners] ** 0.5)`


Neural network configuration
............................

The old stack's `config.training(model=...)` is no longer supported on the new API stack.
Instead, use the new :py:meth:`~ray.rllib.algorithms.algorithm_config.AlgorithmConfig.rl_module`
method to configure RLlib's default :py:class:`~ray.rllib.core.rl_module.rl_module.RLModule`
or specify and configure a custom :py:class:`~ray.rllib.core.rl_module.rl_module.RLModule`.

See :ref:`RLModules API <rlmodule-guide>`, a general guide that also explains
the use of the `config.rl_module()` method.

If you have an old stack `ModelV2` and want to migrate the entire NN logic to the
new stack, see :ref:`ModelV2 to RLModule <rllib-modelv2-to-rlmodule>` for migration instructions.


Learning rate- and coefficient schedules
........................................

If you're using schedules for learning rate or other coefficients, for example, the
`entropy_coeff` setting in PPO, provide scheduling information directly in the respective setting.
Scheduling behavior doesn't require a specific, separate setting anymore.

When defining a schedule, provide a list of 2-tuples, where the first item is the global timestep
(*num_env_steps_sampled_lifetime* in the reported metrics) and the second item is the value that the learning rate should reach at that timestep.
Always start the first 2-tuple with timestep 0. Note that RLlib linearly interpolates values between
two provided timesteps.

For example, to create a learning rate schedule that starts with a value of 1e-5, then increases over 1M timesteps to 1e-4 and stays constant after that, do the following:

.. testcode::

    config.training(
        lr=[
            [0, 1e-5],  # <- initial value at timestep 0
            [1000000, 1e-4],  # <- final value at 1M timesteps
        ],
    )


In the preceding example, the value after 500k timesteps is roughly `5e-5` from linear interpolation.

Another example is to create an entropy coefficient schedule that starts with a value of 0.05, then increases over 1M timesteps to 0.1 and
then suddenly drops to 0, after the 1Mth timestep, do the following:

.. testcode::

    config.training(
        entropy_coeff=[
            [0, 0.05],  # <- initial value at timestep 0
            [1000000, 0.1],  # <- value at 1M timesteps
            [1000001, 0.0],  # <- sudden drop to 0.0 right after 1M timesteps
        ]
    )

In case you need to configure a more complex learning rate scheduling behavior or chain different schedulers
into a pipeline, you can use the experimental `_torch_lr_schedule_classes` config property.
See `this example script <https://github.com/ray-project/ray/blob/master/rllib/examples/learners/ppo_with_torch_lr_schedulers.py>`__  for more details.
Note that this example only covers learning rate schedules, but not any other coefficients.


AlgorithmConfig.learners()

.. testcode::

    # RolloutWorkers have been replace by EnvRunners. EnvRunners are more efficient and offer
    # a more separation-of-concerns design and cleaner code.
    config.env_runners(
        num_env_runners=2,  # use this instead of `num_workers`
    )

    # The following `env_runners` settings are deprecated and should no longer be explicitly
    # set on the new stack:
    config.env_runners(
        create_env_on_local_worker=False,
        sample_collector=None,
        enable_connectors=True,
        remote_worker_envs=False,
        remote_env_batch_wait_ms=0,
        preprocessor_pref="deepmind",
        enable_tf1_exec_eagerly=False,
        sampler_perf_stats_ema_coef=None,
    )

.. hint::

    If you want to IDE-debug what's going on inside your `EnvRunners`, set `num_env_runners=0`
    and make sure you are running your experiment locally and not through Ray Tune.
    In order to do this with any of RLlib's `example <https://github.com/ray-project/ray/tree/master/rllib/examples>`__
    or `tuned_example <https://github.com/ray-project/ray/tree/master/rllib/examples/algorithms>`__ scripts,
    simply set the command line args: `--no-tune --num-env-runners=0`.

In case you were using the `observation_filter` setting, perform the following translations:

.. testcode::

    # For `observation_filter="NoFilter"`, don't set anything in particular. This is the default.

    # For `observation_filter="MeanStdFilter"`, do the following:
    from ray.rllib.connectors.env_to_module import MeanStdFilter

    config.env_runners(
        env_to_module_connector=lambda env: MeanStdFilter(multi_agent=False),  # <- or True
    )


.. hint::
    The main switch for whether to explore or not during sample collection has moved
    to the :py:meth:`~ray.rllib.algorithms.algorithm_config.AlgorithmConfig.env_runners` method.
    See :ref:`here for more details <rllib-algo-config-exploration-docs>`.


.. _rllib-algo-config-exploration-docs:

AlgorithmConfig.exploration()

config.env_runners(explore=True)  # <- or False

# config.callbacks(YourCallbacksClass)

from ray.rllib.callbacks.callbacks import RLlibCallback

class YourCallbacksClass(RLlibCallback):

    def on_episode_start(
        self,
        *,
        episode,
        env_runner,
        metrics_logger,
        env,
        env_index,
        rl_module,

        # Old API stack args; don't use or access these inside your method code.
        worker=None,
        base_env=None,
        policies=None,
        **kwargs,
    ):
        # The `SingleAgentEpisode` or `MultiAgentEpisode` that RLlib has just started.
        # See https://docs.ray.io/en/latest/rllib/single-agent-episode.html for more details:
        print(episode)

        # The `EnvRunner` class that collects the episode in question.
        # This class used to be a `RolloutWorker`. On the new stack, this class is either a
        # `SingleAgentEnvRunner` or a `MultiAgentEnvRunner` holding the gymnasium Env,
        # the RLModule, and the 2 connector pipelines, env-to-module and module-to-env.
        print(env_runner)

        # The MetricsLogger object on the EnvRunner (documentation is a WIP).
        print(metrics_logger.peek("episode_return_mean", default=0.0))

        # The gymnasium env that sample collection uses. Note that this env may be a
        # gymnasium.vector.VectorEnv.
        print(env)

        # The env index, in case of a vector env, that handles the `episode`.
        print(env_index)

        # The RL Module that this EnvRunner uses. Note that this module may be a "plain", single-agent
        # `RLModule`, or a `MultiRLModule` in the multi-agent case.
        print(rl_module)

# Change similarly:
# on_episode_created()
# on_episode_step()
# on_episode_end()

This old API stack method, as well as ``compute_actions`` and ``compute_single_action``, directly translate to
:py:meth:`~ray.rllib.core.rl_module.rl_module.RLModule._forward_inference`
and :py:meth:`~ray.rllib.core.rl_module.rl_module.RLModule._forward_exploration`.
:ref:`The RLModule guide explains how to implement this method <rlmodule-guide>`.


Translating Policy.action_distribution_fn
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To translate ``action_distribution_fn``, write the following custom RLModule code:

.. tab-set::

    .. tab-item:: Same action dist. class

        .. testcode::
            :skipif: True

            from ray.rllib.models.torch.torch_distributions import YOUR_DIST_CLASS


            class MyRLModule(TorchRLModule):
                def setup(self):
                    ...
                    # Set the following attribute at the end of your custom `setup()`.
                    self.action_dist_cls = YOUR_DIST_CLASS


    .. tab-item:: Different action dist. classes

        .. testcode::
            :skipif: True

            from ray.rllib.models.torch.torch_distributions import (
                YOUR_INFERENCE_DIST_CLASS,
                YOUR_EXPLORATION_DIST_CLASS,
                YOUR_TRAIN_DIST_CLASS,
            )

                def get_inference_action_dist_cls(self):
                    return YOUR_INFERENCE_DIST_CLASS

                def get_exploration_action_dist_cls(self):
                    return YOUR_EXPLORATION_DIST_CLASS

                def get_train_action_dist_cls(self):
                    return YOUR_TRAIN_DIST_CLASS


Translating Policy.action_sampler_fn
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To translate ``action_sampler_fn``, write the following custom RLModule code:

.. testcode::
    :skipif: True

    from ray.rllib.models.torch.torch_distributions import YOUR_DIST_CLASS


    class MyRLModule(TorchRLModule):

        def _forward_exploration(self, batch):
            computation_results = ...
            my_dist = YOUR_DIST_CLASS(computation_results)
            actions = my_dist.sample()
            return {Columns.ACTIONS: actions}

        # Maybe for inference, you would like to sample from the deterministic version
        # of your distribution:
        def _forward_inference(self, batch):
            computation_results = ...
            my_dist = YOUR_DIST_CLASS(computation_results)
            greedy_actions = my_dist.to_deterministic().sample()
            return {Columns.ACTIONS: greedy_actions}


Policy.compute_log_likelihoods
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Implement your custom RLModule's :py:meth:`~ray.rllib.core.rl_module.rl_module.RLModule._forward_train` method and
return the ``Columns.ACTION_LOGP`` key together with the corresponding action log probabilities to pass this information
to your loss functions, which your code calls after `forward_train()`. The loss logic can then access
`Columns.ACTION_LOGP`.


Custom loss functions and policies
-------------------------------------

If you're using one or more custom loss functions or custom (PyTorch) optimizers to train your models, instead of doing these
customizations inside the old stack's Policy class, you need to move the logic into the new API stack's
:py:class:`~ray.rllib.core.learner.learner.Learner` class.

See :ref:`Learner <learner-guide>` for details on how to write a custom Learner .

The following example scripts show how to write:

- `a simple custom loss function <https://github.com/ray-project/ray/blob/master/rllib/examples/learners/ppo_with_custom_loss_fn.py>`__
- `a custom Learner with 2 optimizers and different learning rates for each <https://github.com/ray-project/ray/blob/master/rllib/examples/learners/separate_vf_lr_and_optimizer.py>`__.

Note that the new API stack doesn't support the Policy class. In the old stack, this class holds a
neural network, which is the :py:class:`~ray.rllib.core.rl_module.rl_module.RLModule` in the new API stack,
an old stack connector, which is the :py:class:`~ray.rllib.connector.connector_v2.ConnectorV2` in the new API stack,
and one or more optimizers and losses, which are the :py:class:`~ray.rllib.core.learner.learner.Learner` class in the new API stack.

The RL Module API is much more flexible than the old stack's Policy API and
provides a cleaner separation-of-concerns experience. Things related to action
inference run on the EnvRunners, and things related to updating run on the Learner workers
It also provides superior scalability, allowing training in a multi-GPU setup in any Ray cluster
and multi-node with multi-GPU training on the `Anyscale <https://anyscale.com>`__ platform.


Custom connectors
-----------------

If you're using custom connectors from the old API stack, move your logic into the
new :py:class:`~ray.rllib.connectors.connector_v2.ConnectorV2` API.
Translate your agent connectors into env-to-module ConnectorV2 pieces and your
action connectors into module-to-env ConnectorV2 pieces.

The :py:class:`~ray.rllib.connectors.connector_v2.ConnectorV2` documentation is under development.

The following are some examples on how to write ConnectorV2 pieces for the
different pipelines:

#. `Observation frame-stacking <https://github.com/ray-project/ray/blob/master/rllib/examples/connectors/frame_stacking.py>`__.
#. `Add the most recent action and reward to the RL Module's input <https://github.com/ray-project/ray/blob/master/rllib/examples/connectors/prev_actions_prev_rewards.py>`__.
#. `Mean-std filtering on all observations <https://github.com/ray-project/ray/blob/master/rllib/examples/connectors/mean_std_filtering.py>`__.
#. `Flatten any complex observation space to a 1D space <https://github.com/ray-project/ray/blob/master/rllib/examples/connectors/flatten_observations_dict_space.py>`__.