This PR implements a PyTorch version of RLlib's ARS algorithm using RLlib's functional algo builder API. It also adds a regression test for ARS (torch) on CartPole.
SAC (both torch and tf versions) are showing issues (crashes) due to numeric instabilities in the SquashedGaussian distribution (sampling + logp after extreme NN outputs).
This PR fixes these. Stable MuJoCo learning (HalfCheetah) has been confirmed on both tf and torch versions. A Distribution stability test (using extreme NN outputs) has been added for SquashedGaussian (can be used for any other type of distribution as well).
* Policy-classes cleanup and torch/tf unification.
- Make Policy abstract.
- Add `action_dist` to call to `extra_action_out_fn` (necessary for PPO torch).
- Move some methods and vars to base Policy
(from TFPolicy): num_state_tensors, ACTION_PROB, ACTION_LOGP and some more.
* Fix `clip_action` import from Policy (should probably be moved into utils altogether).
* - Move `is_recurrent()` and `num_state_tensors()` into TFPolicy (from DynamicTFPolicy).
- Add config to all Policy c'tor calls (as 3rd arg after obs and action spaces).
* Add `config` to c'tor call to TFPolicy.
* Add missing `config` to c'tor call to TFPolicy in marvil_policy.py.
* Fix test_rollout_worker.py::MockPolicy and BadPolicy classes (Policy base class is now abstract).
* Fix LINT errors in Policy classes.
* Implement StatefulPolicy abstract methods in test cases: test_multi_agent_env.py.
* policy.py LINT errors.
* Create a simple TestPolicy to sub-class from when testing Policies (reduces code in some test cases).
* policy.py
- Remove abstractmethod from `apply_gradients` and `compute_gradients` (these are not required iff `learn_on_batch` implemented).
- Fix docstring of `num_state_tensors`.
* Make QMIX torch Policy a child of TorchPolicy (instead of Policy).
* QMixPolicy add empty implementations of abstract Policy methods.
* Store Policy's config in self.config in base Policy c'tor.
* - Make only compute_actions in base Policy's an abstractmethod and provide pass
implementation to all other methods if not defined.
- Fix state_batches=None (most Policies don't have internal states).
* Cartpole tf learning.
* Cartpole tf AND torch learning (in ~ same ts).
* Cartpole tf AND torch learning (in ~ same ts). 2
* Cartpole tf (torch syntax-broken) learning (in ~ same ts). 3
* Cartpole tf AND torch learning (in ~ same ts). 4
* Cartpole tf AND torch learning (in ~ same ts). 5
* Cartpole tf AND torch learning (in ~ same ts). 6
* Cartpole tf AND torch learning (in ~ same ts). Pendulum tf learning.
* WIP.
* WIP.
* SAC torch learning Pendulum.
* WIP.
* SAC torch and tf learning Pendulum and Cartpole after cleanup.
* WIP.
* LINT.
* LINT.
* SAC: Move policy.target_model to policy.device as well.
* Fixes and cleanup.
* Fix data-format of tf keras Conv2d layers (broken for some tf-versions which have data_format="channels_first" as default).
* Fixes and LINT.
* Fixes and LINT.
* Fix and LINT.
* WIP.
* Test fixes and LINT.
* Fixes and LINT.
Co-authored-by: Sven Mika <sven@Svens-MacBook-Pro.local>
Sven Mika