Add TRPO, D3PG and SAC, minor improvements and bug fixes

README.md

@@ -14,33 +14,36 @@ Welcome to `actorch`, a deep reinforcement learning framework for fast prototypi
 - [REINFORCE](https://people.cs.umass.edu/~barto/courses/cs687/williams92simple.pdf)
 - [Advantage Actor-Critic (A2C)](https://arxiv.org/abs/1602.01783)
 - [Actor-Critic Kronecker-Factored Trust Region (ACKTR)](https://arxiv.org/abs/1708.05144)
+- [Trust Region Policy Optimization (TRPO)](https://arxiv.org/abs/1502.05477)
 - [Proximal Policy Optimization (PPO)](https://arxiv.org/abs/1707.06347)
 - [Advantage-Weighted Regression (AWR)](https://arxiv.org/abs/1910.00177)
 - [Deep Deterministic Policy Gradient (DDPG)](https://arxiv.org/abs/1509.02971)
+- [Distributional Deep Deterministic Policy Gradient (D3PG)](https://arxiv.org/abs/1804.08617)
 - [Twin Delayed Deep Deterministic Policy Gradient (TD3)](https://arxiv.org/abs/1802.09477)
+- [Soft Actor-Critic (SAC)](https://arxiv.org/abs/1801.01290)
 
 ---------------------------------------------------------------------------------------------------------
 
 ## 💡 Key features
 
 - Support for [OpenAI Gymnasium](https://gymnasium.farama.org/) environments
-- Support for custom observation/action spaces
-- Support for custom multimodal input multimodal output models
-- Support for recurrent models (e.g. RNNs, LSTMs, GRUs, etc.)
-- Support for custom policy/value distributions
-- Support for custom preprocessing/postprocessing pipelines
-- Support for custom exploration strategies
+- Support for **custom observation/action spaces**
+- Support for **custom multimodal input multimodal output models**
+- Support for **recurrent models** (e.g. RNNs, LSTMs, GRUs, etc.)
+- Support for **custom policy/value distributions**
+- Support for **custom preprocessing/postprocessing pipelines**
+- Support for **custom exploration strategies**
 - Support for [normalizing flows](https://arxiv.org/abs/1906.02771)
 - Batched environments (both for training and evaluation)
-- Batched trajectory replay
-- Batched and distributional value estimation (e.g. batched and distributional [Retrace](https://arxiv.org/abs/1606.02647) and [V-trace](https://arxiv.org/abs/1802.01561))
-- Data parallel and distributed data parallel multi-GPU training and evaluation
-- Automatic mixed precision training
-- Integration with [Ray Tune](https://docs.ray.io/en/releases-1.13.0/tune/index.html) for experiment execution and hyperparameter tuning at any scale
-- Effortless experiment definition through Python-based configuration files
-- Built-in visualization tool to plot performance metrics
-- Modular object-oriented design
-- Detailed API documentation
+- Batched **trajectory replay**
+- Batched and **distributional value estimation** (e.g. batched and distributional [Retrace](https://arxiv.org/abs/1606.02647) and [V-trace](https://arxiv.org/abs/1802.01561))
+- Data parallel and distributed data parallel **multi-GPU training and evaluation**
+- Automatic **mixed precision training**
+- Integration with [Ray Tune](https://docs.ray.io/en/releases-1.13.0/tune/index.html) for experiment execution and **hyperparameter tuning** at any scale
+- Effortless experiment definition through **Python-based configuration files**
+- Built-in **visualization tool** to plot performance metrics
+- Modular **object-oriented** design
+- Detailed **API documentation**
 
 ---------------------------------------------------------------------------------------------------------
 
@@ -161,7 +164,7 @@ experiment_params = ExperimentParams(
         enable_amp=False,
         enable_reproducibility=True,
         log_sys_usage=True,
-        suppress_warnings=False,
+        suppress_warnings=True,
     ),
 )
 ```
@@ -197,6 +200,8 @@ You can find the generated plots in `plots`.
 
 Congratulations, you ran your first experiment!
 
+See `examples` for additional configuration file examples.
+
 **HINT**: since a configuration file is a regular Python script, you can use all the
 features of the language (e.g. inheritance).
 
@@ -217,6 +222,21 @@ features of the language (e.g. inheritance).
 
 ---------------------------------------------------------------------------------------------------------
 
+## @ Citation
+
+```
+@misc{DellaLibera2022ACTorch,
+  author = {Luca Della Libera},
+  title = {{ACTorch}: a Deep Reinforcement Learning Framework for Fast Prototyping},
+  year = {2022},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/lucadellalib/actorch}},
+}
+```
+
+---------------------------------------------------------------------------------------------------------
+
 ## 📧 Contact
 
 [luca.dellalib@gmail.com](mailto:luca.dellalib@gmail.com)

actorch/algorithms/__init__.py

@@ -20,7 +20,10 @@
 from actorch.algorithms.acktr import *
 from actorch.algorithms.algorithm import *
 from actorch.algorithms.awr import *
+from actorch.algorithms.d3pg import *
 from actorch.algorithms.ddpg import *
 from actorch.algorithms.ppo import *
 from actorch.algorithms.reinforce import *
+from actorch.algorithms.sac import *
 from actorch.algorithms.td3 import *
+from actorch.algorithms.trpo import *

actorch/algorithms/a2c.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 # ==============================================================================
 
-"""Advantage Actor-Critic."""
+"""Advantage Actor-Critic (A2C)."""
 
 import contextlib
 from typing import Any, Callable, Dict, Optional, Tuple, Type, Union
@@ -38,7 +38,7 @@
     DistributedDataParallelREINFORCE,
     LRScheduler,
 )
-from actorch.algorithms.utils import prepare_model
+from actorch.algorithms.utils import normalize_, prepare_model
 from actorch.algorithms.value_estimation import n_step_return
 from actorch.distributions import Deterministic
 from actorch.envs import BatchedEnv
@@ -65,7 +65,7 @@
 
 
 class A2C(REINFORCE):
-    """Advantage Actor-Critic.
+    """Advantage Actor-Critic (A2C).
 
     References
     ----------
@@ -208,12 +208,8 @@ def setup(self, config: "Dict[str, Any]") -> "None":
         self.config = A2C.Config(**self.config)
         self.config["_accept_kwargs"] = True
         super().setup(config)
-        self._value_network = (
-            self._build_value_network().train().to(self._device, non_blocking=True)
-        )
-        self._value_network_loss = (
-            self._build_value_network_loss().train().to(self._device, non_blocking=True)
-        )
+        self._value_network = self._build_value_network()
+        self._value_network_loss = self._build_value_network_loss()
         self._value_network_optimizer = self._build_value_network_optimizer()
         self._value_network_optimizer_lr_scheduler = (
             self._build_value_network_optimizer_lr_scheduler()
@@ -324,16 +320,20 @@ def _build_value_network(self) -> "Network":
             self.value_network_normalizing_flows,
         )
         self._log_graph(value_network.wrapped_model.model, "value_network_model")
-        return value_network
+        return value_network.train().to(self._device, non_blocking=True)
 
     def _build_value_network_loss(self) -> "Loss":
         if self.value_network_loss_builder is None:
             self.value_network_loss_builder = torch.nn.MSELoss
         if self.value_network_loss_config is None:
             self.value_network_loss_config: "Dict[str, Any]" = {}
-        return self.value_network_loss_builder(
-            reduction="none",
-            **self.value_network_loss_config,
+        return (
+            self.value_network_loss_builder(
+                reduction="none",
+                **self.value_network_loss_config,
+            )
+            .train()
+            .to(self._device, non_blocking=True)
         )
 
     def _build_value_network_optimizer(self) -> "Optimizer":
@@ -374,6 +374,8 @@ def _train_step(self) -> "Dict[str, Any]":
         result = super()._train_step()
         self.num_return_steps.step()
         result["num_return_steps"] = self.num_return_steps()
+        result["entropy_coeff"] = result.pop("entropy_coeff", None)
+        result["max_grad_l2_norm"] = result.pop("max_grad_l2_norm", None)
         return result
 
     # override
@@ -405,17 +407,7 @@ def _train_on_batch(
                     self.num_return_steps(),
                 )
             if self.normalize_advantage:
-                length = mask.sum(dim=1, keepdim=True)
-                advantages_mean = advantages.sum(dim=1, keepdim=True) / length
-                advantages -= advantages_mean
-                advantages *= mask
-                advantages_stddev = (
-                    ((advantages**2).sum(dim=1, keepdim=True) / length)
-                    .sqrt()
-                    .clamp(min=1e-6)
-                )
-                advantages /= advantages_stddev
-                advantages *= mask
+                normalize_(advantages, dim=-1, mask=mask)
 
         # Discard next state value
         state_values = state_values[:, :-1]
@@ -449,10 +441,12 @@ def _train_on_batch_value_network(
             state_value = state_values[mask]
             target = targets[mask]
             loss = self._value_network_loss(state_value, target)
-            loss *= is_weight[:, None].expand_as(mask)[mask]
+            priority = None
+            if self._buffer.is_prioritized:
+                loss *= is_weight[:, None].expand_as(mask)[mask]
+                priority = loss.detach().abs().to("cpu").numpy()
             loss = loss.mean()
         optimize_result = self._optimize_value_network(loss)
-        priority = None
         result = {
             "state_value": state_value.mean().item(),
             "target": target.mean().item(),
@@ -490,7 +484,7 @@ def _get_default_value_network_preprocessor(
 
 
 class DistributedDataParallelA2C(DistributedDataParallelREINFORCE):
-    """Distributed data parallel Advantage Actor-Critic.
+    """Distributed data parallel Advantage Actor-Critic (A2C).
 
     See Also
     --------

actorch/algorithms/acktr.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 # ==============================================================================
 
-"""Actor-Critic Kronecker-Factored Trust Region."""
+"""Actor-Critic Kronecker-Factored Trust Region (ACKTR)."""
 
 from typing import Any, Callable, Dict, Optional, Union
 
@@ -43,7 +43,7 @@
 
 
 class ACKTR(A2C):
-    """Actor-Critic Kronecker-Factored Trust Region.
+    """Actor-Critic Kronecker-Factored Trust Region (ACKTR).
 
     References
     ----------
@@ -287,7 +287,7 @@ def _optimize_policy_network(self, loss: "Tensor") -> "Dict[str, Any]":
 
 
 class DistributedDataParallelACKTR(DistributedDataParallelA2C):
-    """Distributed data parallel Actor-Critic Kronecker-Factored Trust Region.
+    """Distributed data parallel Actor-Critic Kronecker-Factored Trust Region (ACKTR).
 
     See Also
     --------

actorch/algorithms/algorithm.py

@@ -51,7 +51,6 @@
 from ray.tune.syncer import NodeSyncer
 from ray.tune.trial import ExportFormat
 from torch import Tensor
-from torch.cuda.amp import autocast
 from torch.distributions import Bernoulli, Categorical, Distribution, Normal
 from torch.profiler import profile, record_function, tensorboard_trace_handler
 from torch.utils.data import DataLoader
@@ -113,7 +112,7 @@ class Algorithm(ABC, Trainable):
 
     _EXPORT_FORMATS = [ExportFormat.CHECKPOINT, ExportFormat.MODEL]
 
-    _UPDATE_BUFFER_DATASET_SCHEDULES_AFTER_TRAIN_EPOCH = True
+    _OFF_POLICY = True
 
     class Config(dict):
         """Keyword arguments expected in the configuration received by `setup`."""
@@ -692,12 +691,7 @@ def _build_train_env(self) -> "BatchedEnv":
         if self.train_env_config is None:
             self.train_env_config = {}
 
-        try:
-            train_env = self.train_env_builder(
-                **self.train_env_config,
-            )
-        except TypeError:
-            train_env = self.train_env_builder(**self.train_env_config)
+        train_env = self.train_env_builder(**self.train_env_config)
         if not isinstance(train_env, BatchedEnv):
             train_env.close()
             train_env = SerialBatchedEnv(self.train_env_builder, self.train_env_config)
@@ -866,7 +860,7 @@ def _build_policy_network(self) -> "PolicyNetwork":  # noqa: C901
             self.policy_network_postprocessors,
         )
         self._log_graph(policy_network.wrapped_model.model, "policy_network_model")
-        return policy_network
+        return policy_network.train().to(self._device, non_blocking=True)
 
     def _build_train_agent(self) -> "Agent":
         if self.train_agent_builder is None:
@@ -989,14 +983,18 @@ def _build_dataloader(self) -> "DataLoader":
         if self.dataloader_builder is None:
             self.dataloader_builder = DataLoader
             if self.dataloader_config is None:
-                fork = torch.multiprocessing.get_start_method() == "fork"
+                use_mp = (
+                    self._OFF_POLICY
+                    and not self._buffer.is_prioritized
+                    and torch.multiprocessing.get_start_method() == "fork"
+                )
                 self.dataloader_config = {
-                    "num_workers": 1 if fork else 0,
+                    "num_workers": 1 if use_mp else 0,
                     "pin_memory": True,
                     "timeout": 0,
                     "worker_init_fn": None,
                     "generator": None,
-                    "prefetch_factor": 1 if fork else 2,
+                    "prefetch_factor": 1 if use_mp else 2,
                     "pin_memory_device": "",
                 }
         if self.dataloader_config is None:
@@ -1037,20 +1035,15 @@ def _train_step(self) -> "Dict[str, Any]":
         if self.train_num_episodes_per_iter:
             train_num_episodes_per_iter = self.train_num_episodes_per_iter()
             self.train_num_episodes_per_iter.step()
-        with (
-            autocast(**self.enable_amp)
-            if self.enable_amp["enabled"]
-            else contextlib.suppress()
+        for experience, done in self._train_sampler.sample(
+            train_num_timesteps_per_iter,
+            train_num_episodes_per_iter,
         ):
-            for experience, done in self._train_sampler.sample(
-                train_num_timesteps_per_iter,
-                train_num_episodes_per_iter,
-            ):
-                self._buffer.add(experience, done)
+            self._buffer.add(experience, done)
         result = self._train_sampler.stats
         self._cumrewards += result["episode_cumreward"]
 
-        if not self._UPDATE_BUFFER_DATASET_SCHEDULES_AFTER_TRAIN_EPOCH:
+        if not self._OFF_POLICY:
             for schedule in self._buffer_dataset.schedules.values():
                 schedule.step()
         train_epoch_result = self._train_epoch()
@@ -1060,7 +1053,7 @@ def _train_step(self) -> "Dict[str, Any]":
             schedule.step()
         for schedule in self._buffer.schedules.values():
             schedule.step()
-        if self._UPDATE_BUFFER_DATASET_SCHEDULES_AFTER_TRAIN_EPOCH:
+        if self._OFF_POLICY:
             for schedule in self._buffer_dataset.schedules.values():
                 schedule.step()
         return result
@@ -1113,16 +1106,11 @@ def _eval_step(self) -> "Dict[str, Any]":
             eval_num_episodes_per_iter = self.eval_num_episodes_per_iter()
             self.eval_num_episodes_per_iter.step()
         self._eval_sampler.reset()
-        with (
-            autocast(**self.enable_amp)
-            if self.enable_amp["enabled"]
-            else contextlib.suppress()
+        for _ in self._eval_sampler.sample(
+            eval_num_timesteps_per_iter,
+            eval_num_episodes_per_iter,
         ):
-            for _ in self._eval_sampler.sample(
-                eval_num_timesteps_per_iter,
-                eval_num_episodes_per_iter,
-            ):
-                pass
+            pass
         for schedule in self._eval_agent.schedules.values():
             schedule.step()
         return self._eval_sampler.stats
@@ -1353,7 +1341,7 @@ def __init__(
                 Default to ``{}``.
             placement_strategy:
                 The placement strategy
-                (see https://docs.ray.io/en/latest/ray-core/placement-group.html).
+                (see https://docs.ray.io/en/releases-1.13.0/ray-core/placement-group.html for Ray 1.13.0).
             backend:
                 The backend for distributed execution
                 (see https://pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group).