Add Pytorch TRPO

README.md

@@ -43,7 +43,7 @@ The table below summarizes the algorithms available in garage.
 | REPS                   | TensorFlow          |
 | TD3                    | TensorFlow          |
 | TNPG                   | TensorFlow          |
-| TRPO                   | TensorFlow          |
+| TRPO                   | PyTorch, TensorFlow |
 
 ## Supported Tools and Frameworks
 garage supports Python 3.5+.

examples/torch/trpo_pendulum.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+"""This is an example to train a task with TRPO algorithm (PyTorch).
+
+Here it runs InvertedDoublePendulum-v2 environment with 100 iterations.
+"""
+import torch
+
+from garage.experiment import LocalRunner, run_experiment
+from garage.np.baselines import LinearFeatureBaseline
+from garage.tf.envs import TfEnv
+from garage.torch.algos import TRPO
+from garage.torch.policies import GaussianMLPPolicy
+
+
+def run_task(snapshot_config, *_):
+    """Set up environment and algorithm and run the task.
+
+    Args:
+        snapshot_config (garage.experiment.SnapshotConfig): The snapshot
+            configuration used by LocalRunner to create the snapshotter.
+            If None, it will create one with default settings.
+        _ : Unused parameters
+
+    """
+    env = TfEnv(env_name='InvertedDoublePendulum-v2')
+
+    runner = LocalRunner(snapshot_config)
+
+    policy = GaussianMLPPolicy(env.spec,
+                               hidden_sizes=[32, 32],
+                               hidden_nonlinearity=torch.tanh,
+                               output_nonlinearity=None)
+
+    baseline = LinearFeatureBaseline(env_spec=env.spec)
+
+    algo = TRPO(env_spec=env.spec,
+                policy=policy,
+                baseline=baseline,
+                max_path_length=100,
+                discount=0.99,
+                center_adv=False)
+
+    runner.setup(algo, env)
+    runner.train(n_epochs=100, batch_size=1024)
+
+
+run_experiment(
+    run_task,
+    snapshot_mode='last',
+    seed=1,
+)

src/garage/torch/algos/__init__.py

@@ -2,5 +2,6 @@
 from garage.torch.algos.ddpg import DDPG
 from garage.torch.algos.vpg import VPG
 from garage.torch.algos.ppo import PPO  # noqa: I100
+from garage.torch.algos.trpo import TRPO
 
-__all__ = ['DDPG', 'VPG', 'PPO']
+__all__ = ['DDPG', 'VPG', 'PPO', 'TRPO']

src/garage/torch/algos/ppo.py

@@ -1,6 +1,4 @@
 """Proximal Policy Optimization (PPO)."""
-import copy
-
 import torch
 
 from garage.torch.algos import VPG
@@ -29,6 +27,7 @@ class PPO(VPG):
             standardized before shifting.
         optimizer (object): The optimizer of the algorithm. Should be the
             optimizers in torch.optim.
+        optimizer_args (dict): Arguments required to initialize the optimizer.
         policy_ent_coeff (float): The coefficient of the policy entropy.
             Setting it to zero would mean no entropy regularization.
         use_softplus_entropy (bool): Whether to estimate the softmax
@@ -56,18 +55,18 @@ def __init__(self,
                  center_adv=True,
                  positive_adv=False,
                  optimizer=None,
+                 optimizer_args=None,
                  policy_ent_coeff=0.0,
                  use_softplus_entropy=False,
                  stop_entropy_gradient=False,
                  entropy_method='no_entropy'):
         super().__init__(env_spec, policy, baseline, max_path_length,
                          policy_lr, n_samples, discount, gae_lambda,
-                         center_adv, positive_adv, optimizer, policy_ent_coeff,
-                         use_softplus_entropy, stop_entropy_gradient,
-                         entropy_method)
+                         center_adv, positive_adv, optimizer, optimizer_args,
+                         policy_ent_coeff, use_softplus_entropy,
+                         stop_entropy_gradient, entropy_method)
 
         self._lr_clip_range = lr_clip_range
-        self._old_policy = copy.deepcopy(self.policy)
 
     def _compute_objective(self, advantages, valids, obs, actions, rewards):
         """Compute objective using surrogate value and clipped surrogate value.
@@ -83,17 +82,13 @@ def _compute_objective(self, advantages, valids, obs, actions, rewards):
             torch.Tensor: Calculated objective values
 
         """
-        # pylint: disable=unused-argument
         # Compute constraint
         with torch.no_grad():
             old_ll = self._old_policy.log_likelihood(obs, actions)
         new_ll = self.policy.log_likelihood(obs, actions)
 
         likelihood_ratio = (new_ll - old_ll).exp()
 
-        # Memorize the policy state_dict
-        self._old_policy.load_state_dict(self.policy.state_dict())
-
         # Calculate surrogate
         surrogate = likelihood_ratio * advantages
 

src/garage/torch/algos/trpo.py

@@ -0,0 +1,104 @@
+"""Trust Region Policy Optimization."""
+import torch
+
+from garage.torch.algos import VPG
+from garage.torch.optimizers import ConjugateGradientOptimizer
+
+
+class TRPO(VPG):
+    """Trust Region Policy Optimization (TRPO).
+
+    Args:
+        env_spec (garage.envs.EnvSpec): Environment specification.
+        policy (garage.torch.policies.base.Policy): Policy.
+        baseline (garage.np.baselines.Baseline): The baseline.
+        max_path_length (int): Maximum length of a single rollout.
+        policy_lr (float): Learning rate for training policy network.
+        num_train_per_epoch (int): Number of train_once calls per epoch.
+        discount (float): Discount.
+        gae_lambda (float): Lambda used for generalized advantage
+            estimation.
+        max_kl (float): Maximum KL divergence between old and new policies.
+        center_adv (bool): Whether to rescale the advantages
+            so that they have mean 0 and standard deviation 1.
+        positive_adv (bool): Whether to shift the advantages
+            so that they are always positive. When used in
+            conjunction with center_adv the advantages will be
+            standardized before shifting.
+        optimizer (object): The optimizer of the algorithm. Should be the
+            optimizers in torch.optim.
+        optimizer_args (dict): Arguments required to initialize the optimizer.
+        policy_ent_coeff (float): The coefficient of the policy entropy.
+            Setting it to zero would mean no entropy regularization.
+        use_softplus_entropy (bool): Whether to estimate the softmax
+            distribution of the entropy to prevent the entropy from being
+            negative.
+        stop_entropy_gradient (bool): Whether to stop the entropy gradient.
+        entropy_method (str): A string from: 'max', 'regularized',
+            'no_entropy'. The type of entropy method to use. 'max' adds the
+            dense entropy to the reward for each time step. 'regularized' adds
+            the mean entropy to the surrogate objective. See
+            https://arxiv.org/abs/1805.00909 for more details.
+
+    """
+
+    def __init__(self,
+                 env_spec,
+                 policy,
+                 baseline,
+                 max_path_length=100,
+                 policy_lr=3e-4,
+                 num_train_per_epoch=1,
+                 discount=0.99,
+                 gae_lambda=0.98,
+                 max_kl=0.01,
+                 center_adv=True,
+                 positive_adv=False,
+                 optimizer=None,
+                 optimizer_args=None,
+                 policy_ent_coeff=0.0,
+                 use_softplus_entropy=False,
+                 stop_entropy_gradient=False,
+                 entropy_method='no_entropy'):
+        if optimizer is None:
+            optimizer = ConjugateGradientOptimizer
+            optimizer_args = {'max_constraint_value': max_kl}
+
+        super().__init__(env_spec, policy, baseline, max_path_length,
+                         policy_lr, num_train_per_epoch, discount, gae_lambda,
+                         center_adv, positive_adv, optimizer, optimizer_args,
+                         policy_ent_coeff, use_softplus_entropy,
+                         stop_entropy_gradient, entropy_method)
+
+        self._kl = None
+
+    def _compute_objective(self, advantages, valids, obs, actions, rewards):
+        """Compute the surrogate objective.
+
+        Args:
+            advantages (torch.Tensor): Expected rewards over the actions
+            valids (list[int]): length of the valid values for each path
+            obs (torch.Tensor): Observation from the environment.
+            actions (torch.Tensor): Predicted action.
+            rewards (torch.Tensor): Feedback from the environment.
+
+        Returns:
+            torch.Tensor: Calculated objective values
+
+        """
+        with torch.no_grad():
+            old_ll = self._old_policy.log_likelihood(obs, actions)
+
+        new_ll = self.policy.log_likelihood(obs, actions)
+        likelihood_ratio = (new_ll - old_ll).exp()
+
+        # Calculate surrogate
+        surrogate = likelihood_ratio * advantages
+
+        return surrogate
+
+    def _optimize(self, itr, paths, valids, obs, actions, rewards):
+        self._optimizer.step(
+            f_loss=lambda: self._compute_loss(itr, paths, valids, obs, actions,
+                                              rewards),
+            f_constraint=lambda: self._compute_kl_constraint(obs))

src/garage/torch/algos/vpg.py

@@ -1,5 +1,6 @@
 """Vanilla Policy Gradient (REINFORCE)."""
 import collections
+import copy
 
 from dowel import tabular
 import numpy as np
@@ -9,6 +10,7 @@
 from garage.misc import tensor_utils
 from garage.np.algos import BatchPolopt
 from garage.torch.algos import loss_function_utils
+from garage.torch.utils import flatten_batch
 
 
 class VPG(BatchPolopt):
@@ -34,6 +36,7 @@ class VPG(BatchPolopt):
             standardized before shifting.
         optimizer (object): The optimizer of the algorithm. Should be the
             optimizers in torch.optim.
+        optimizer_args (dict): Arguments required to initialize the optimizer.
         policy_ent_coeff (float): The coefficient of the policy entropy.
             Setting it to zero would mean no entropy regularization.
         use_softplus_entropy (bool): Whether to estimate the softmax
@@ -61,6 +64,7 @@ def __init__(
             center_adv=True,
             positive_adv=False,
             optimizer=None,
+            optimizer_args=None,
             policy_ent_coeff=0.0,
             use_softplus_entropy=False,
             stop_entropy_gradient=False,
@@ -83,16 +87,20 @@ def __init__(
                                           stop_entropy_gradient,
                                           policy_ent_coeff)
         self._episode_reward_mean = collections.deque(maxlen=100)
-        self._optimizer = optimizer(policy.parameters(),
-                                    lr=policy_lr,
-                                    eps=1e-5)
+
+        if optimizer_args is None:
+            optimizer_args = {'lr': policy_lr, 'eps': 1e-5}
+
+        self._optimizer = optimizer(policy.parameters(), **optimizer_args)
 
         super().__init__(policy=policy,
                          baseline=baseline,
                          discount=discount,
                          max_path_length=max_path_length,
                          n_samples=n_samples)
 
+        self._old_policy = copy.deepcopy(self.policy)
+
     @staticmethod
     def _check_entropy_configuration(entropy_method, center_adv,
                                      stop_entropy_gradient, policy_ent_coeff):
@@ -124,15 +132,27 @@ def train_once(self, itr, paths):
 
         """
         valids, obs, actions, rewards = self.process_samples(itr, paths)
-        average_return = self._log(itr, paths)
 
         loss = self._compute_loss(itr, paths, valids, obs, actions, rewards)
 
+        # Memorize the policy state_dict
+        self._old_policy.load_state_dict(self.policy.state_dict())
+
         self._optimizer.zero_grad()
-        # using a negative because optimizers use gradient descent,
-        # whilst we want gradient ascent.
-        (-loss).backward()
-        self._optimizer.step()
+        loss.backward()
+
+        kl_before = self._compute_kl_constraint(obs).detach()
+        self._optimize(itr, paths, valids, obs, actions, rewards)
+
+        with torch.no_grad():
+            loss_after = self._compute_loss(itr, paths, valids, obs, actions,
+                                            rewards)
+            kl = self._compute_kl_constraint(obs)
+            policy_entropy = self._compute_policy_entropy(obs)
+            average_return = self._log(itr, paths, loss.item(),
+                                       loss_after.item(), kl_before.item(),
+                                       kl.item(),
+                                       policy_entropy.mean().item())
 
         self.baseline.fit(paths)
         return average_return
@@ -188,7 +208,31 @@ def _compute_loss(self, itr, paths, valids, obs, actions, rewards):
             objective += self._policy_ent_coeff * policy_entropies
 
         valid_objectives = loss_function_utils.filter_valids(objective, valids)
-        return torch.cat(valid_objectives).mean()
+        return -torch.cat(valid_objectives).mean()
+
+    def _compute_kl_constraint(self, obs):
+        """Compute KL divergence.
+
+        Compute the KL divergence between the old policy distribution and
+        current policy distribution.
+
+        Args:
+            obs (torch.Tensor): Observation from the environment.
+
+        Returns:
+            torch.Tensor: Calculated mean KL divergence.
+
+        """
+        flat_obs = flatten_batch(obs)
+        with torch.no_grad():
+            old_dist = self._old_policy.forward(flat_obs)
+
+        new_dist = self.policy.forward(flat_obs)
+
+        kl_constraint = torch.distributions.kl.kl_divergence(
+            old_dist, new_dist)
+
+        return kl_constraint.mean()
 
     def _compute_policy_entropy(self, obs):
         """Compute entropy value of probability distribution.
@@ -247,6 +291,9 @@ def _get_baselines(self, path):
             return torch.Tensor(self.baseline.predict_n(path))
         return torch.Tensor(self.baseline.predict(path))
 
+    def _optimize(self, itr, paths, valids, obs, actions, rewards):  # pylint: disable=unused-argument  # noqa: E501
+        self._optimizer.step()
+
     def process_samples(self, itr, paths):
         """Process sample data based on the collected paths.
 
@@ -282,12 +329,18 @@ def process_samples(self, itr, paths):
 
         return valids, obs, actions, rewards
 
-    def _log(self, itr, paths):
+    def _log(self, itr, paths, loss_before, loss_after, kl_before, kl,
+             policy_entropy):
         """Log information per iteration based on the collected paths.
 
         Args:
             itr (int): Iteration number.
             paths (list[dict]): A list of collected paths
+            loss_before (float): Loss before optimization step.
+            loss_after (float): Loss after optimization step.
+            kl_before (float): KL divergence before optimization step.
+            kl (float): KL divergence after optimization step.
+            policy_entropy (float): Policy entropy.
 
         Returns:
             float: The average return in last epoch cycle.
@@ -308,5 +361,12 @@ def _log(self, itr, paths):
         tabular.record('StdReturn', np.std(undiscounted_returns))
         tabular.record('MaxReturn', np.max(undiscounted_returns))
         tabular.record('MinReturn', np.min(undiscounted_returns))
+        with tabular.prefix(self.policy.name):
+            tabular.record('LossBefore', loss_before)
+            tabular.record('LossAfter', loss_after)
+            tabular.record('dLoss', loss_before - loss_after)
+            tabular.record('KLBefore', kl_before)
+            tabular.record('KL', kl)
+            tabular.record('Entropy', policy_entropy)
 
         return average_return

src/garage/torch/optimizers/__init__.py

@@ -0,0 +1,5 @@
+"""PyTorch optimizers."""
+from garage.torch.optimizers.conjugate_gradient_optimizer import (
+    ConjugateGradientOptimizer)
+
+__all__ = ['ConjugateGradientOptimizer']