# Copyright 2020,2021 Sony Corporation.
# Copyright 2021,2022,2023 Sony Group Corporation.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from typing import Any, Dict, List, Union
import gym
import nnabla as nn
import nnabla.solvers as NS
import nnabla_rl.environment_explorers as EE
import nnabla_rl.model_trainers as MT
from nnabla_rl.algorithm import Algorithm, AlgorithmConfig, eval_api
from nnabla_rl.algorithms.common_utils import _StochasticPolicyActionSelector
from nnabla_rl.builders import ExplorerBuilder, ModelBuilder, ReplayBufferBuilder, SolverBuilder
from nnabla_rl.environment_explorer import EnvironmentExplorer
from nnabla_rl.environments.environment_info import EnvironmentInfo
from nnabla_rl.model_trainers.model_trainer import ModelTrainer, TrainingBatch
from nnabla_rl.models import QFunction, SACPolicy, SACQFunction, SACVFunction, StochasticPolicy, VFunction
from nnabla_rl.replay_buffer import ReplayBuffer
from nnabla_rl.utils import context
from nnabla_rl.utils.data import marshal_experiences
from nnabla_rl.utils.misc import sync_model
[docs]@dataclass
class ICML2018SACConfig(AlgorithmConfig):
"""ICML2018SACConfig List of configurations for ICML2018SAC algorithm.
Args:
gamma (float): discount factor of rewards. Defaults to 0.99.
learning_rate (float): learning rate which is set to all solvers. \
You can customize/override the learning rate for each solver by implementing the \
(:py:class:`SolverBuilder <nnabla_rl.builders.SolverBuilder>`) by yourself. \
Defaults to 0.0003.
batch_size(int): training batch size. Defaults to 256.
tau (float): target network's parameter update coefficient. Defaults to 0.005.
environment_steps (int): Number of steps to interact with the environment on each iteration. Defaults to 1.
gradient_steps (int): Number of parameter updates to perform on each iteration. Defaults to 1.
reward_scalar (float): Reward scaling factor. Obtained reward will be multiplied by this value. Defaults to 5.0.
start_timesteps (int): the timestep when training starts.\
The algorithm will collect experiences from the environment by acting randomly until this timestep.\
Defaults to 10000.
replay_buffer_size (int): capacity of the replay buffer. Defaults to 1000000.
num_steps (int): number of steps for N-step Q targets. Defaults to 1.
target_update_interval (float): the interval of target v function parameter's update. Defaults to 1.
pi_unroll_steps (int): Number of steps to unroll policy's tranining network.\
The network will be unrolled even though the provided model doesn't have RNN layers.\
Defaults to 1.
pi_burn_in_steps (int): Number of burn-in steps to initiaze policy's recurrent layer states during training.\
This flag does not take effect if given model is not an RNN model.\
Defaults to 0.
pi_reset_rnn_on_terminal (bool): Reset policy's recurrent internal states to zero during training\
if episode ends. This flag does not take effect if given model is not an RNN model.\
Defaults to False.
q_unroll_steps (int): Number of steps to unroll q-function's tranining network.\
The network will be unrolled even though the provided model doesn't have RNN layers.\
Defaults to 1.
q_burn_in_steps (int): Number of burn-in steps to initiaze q-function's recurrent layer states\
during training. This flag does not take effect if given model is not an RNN model.\
Defaults to 0.
q_reset_rnn_on_terminal (bool): Reset q-function's recurrent internal states to zero during training\
if episode ends. This flag does not take effect if given model is not an RNN model.\
Defaults to False.
v_unroll_steps (int): Number of steps to unroll v-function's tranining network.\
The network will be unrolled even though the provided model doesn't have RNN layers.\
Defaults to 1.
v_burn_in_steps (int): Number of burn-in steps to initiaze v-function's recurrent layer states\
during training. This flag does not take effect if given model is not an RNN model.\
Defaults to 0.
v_reset_rnn_on_terminal (bool): Reset v-function's recurrent internal states to zero during training\
if episode ends. This flag does not take effect if given model is not an RNN model.\
Defaults to False.
"""
gamma: float = 0.99
learning_rate: float = 3.0*1e-4
batch_size: int = 256
tau: float = 0.005
environment_steps: int = 1
gradient_steps: int = 1
reward_scalar: float = 5.0
start_timesteps: int = 10000
replay_buffer_size: int = 1000000
target_update_interval: int = 1
num_steps: int = 1
# rnn model support
pi_unroll_steps: int = 1
pi_burn_in_steps: int = 0
pi_reset_rnn_on_terminal: bool = True
q_unroll_steps: int = 1
q_burn_in_steps: int = 0
q_reset_rnn_on_terminal: bool = True
v_unroll_steps: int = 1
v_burn_in_steps: int = 0
v_reset_rnn_on_terminal: bool = True
def __post_init__(self):
"""__post_init__
Check the values are in valid range.
"""
self._assert_between(self.tau, 0.0, 1.0, 'tau')
self._assert_between(self.gamma, 0.0, 1.0, 'gamma')
self._assert_positive(self.gradient_steps, 'gradient_steps')
self._assert_positive(self.environment_steps, 'environment_steps')
self._assert_positive(self.start_timesteps, 'start_timesteps')
self._assert_positive(self.target_update_interval, 'target_update_interval')
self._assert_positive(self.num_steps, 'num_steps')
self._assert_positive(self.pi_unroll_steps, 'pi_unroll_steps')
self._assert_positive_or_zero(self.pi_burn_in_steps, 'pi_burn_in_steps')
self._assert_positive(self.q_unroll_steps, 'q_unroll_steps')
self._assert_positive_or_zero(self.q_burn_in_steps, 'q_burn_in_steps')
self._assert_positive(self.v_unroll_steps, 'v_unroll_steps')
self._assert_positive_or_zero(self.v_burn_in_steps, 'v_burn_in_steps')
class DefaultVFunctionBuilder(ModelBuilder[VFunction]):
def build_model(self, # type: ignore[override]
scope_name: str,
env_info: EnvironmentInfo,
algorithm_config: ICML2018SACConfig,
**kwargs) -> VFunction:
return SACVFunction(scope_name)
class DefaultQFunctionBuilder(ModelBuilder[QFunction]):
def build_model(self, # type: ignore[override]
scope_name: str,
env_info: EnvironmentInfo,
algorithm_config: ICML2018SACConfig,
**kwargs) -> QFunction:
return SACQFunction(scope_name)
class DefaultPolicyBuilder(ModelBuilder[StochasticPolicy]):
def build_model(self, # type: ignore[override]
scope_name: str,
env_info: EnvironmentInfo,
algorithm_config: ICML2018SACConfig,
**kwargs) -> StochasticPolicy:
return SACPolicy(scope_name, env_info.action_dim)
class DefaultSolverBuilder(SolverBuilder):
def build_solver(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: ICML2018SACConfig,
**kwargs) -> nn.solver.Solver:
assert isinstance(algorithm_config, ICML2018SACConfig)
return NS.Adam(alpha=algorithm_config.learning_rate)
class DefaultReplayBufferBuilder(ReplayBufferBuilder):
def build_replay_buffer(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: ICML2018SACConfig,
**kwargs) -> ReplayBuffer:
assert isinstance(algorithm_config, ICML2018SACConfig)
return ReplayBuffer(capacity=algorithm_config.replay_buffer_size)
class DefaultExplorerBuilder(ExplorerBuilder):
def build_explorer(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: ICML2018SACConfig,
algorithm: "ICML2018SAC",
**kwargs) -> EnvironmentExplorer:
explorer_config = EE.RawPolicyExplorerConfig(
warmup_random_steps=algorithm_config.start_timesteps,
reward_scalar=algorithm_config.reward_scalar,
initial_step_num=algorithm.iteration_num,
timelimit_as_terminal=False
)
explorer = EE.RawPolicyExplorer(policy_action_selector=algorithm._exploration_action_selector,
env_info=env_info,
config=explorer_config)
return explorer
[docs]class ICML2018SAC(Algorithm):
"""Soft Actor-Critic (SAC) algorithm.
This class implements the ICML2018 version of Soft Actor Critic (SAC) algorithm proposed by T. Haarnoja, et al.
in the paper: "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor"
For detail see: https://arxiv.org/abs/1801.01290
This implementation slightly differs from the implementation of Soft Actor-Critic algorithm presented
also by T. Haarnoja, et al. in the following paper: https://arxiv.org/abs/1812.05905
You will need to scale the reward received from the environment properly to get the algorithm work.
Args:
env_or_env_info \
(gym.Env or :py:class:`EnvironmentInfo <nnabla_rl.environments.environment_info.EnvironmentInfo>`):
the environment to train or environment info
config (:py:class:`ICML2018SACConfig <nnabla_rl.algorithms.icml2018_sac.ICML2018SACConfig>`):
configuration of the ICML2018SAC algorithm
v_function_builder (:py:class:`ModelBuilder[VFunction] <nnabla_rl.builders.ModelBuilder>`):
builder of v function models
v_solver_builder (:py:class:`SolverBuilder <nnabla_rl.builders.SolverBuilder>`):
builder of v function solvers
q_function_builder (:py:class:`ModelBuilder[QFunction] <nnabla_rl.builders.ModelBuilder>`):
builder of q function models
q_solver_builder (:py:class:`SolverBuilder <nnabla_rl.builders.SolverBuilder>`):
builder of q function solvers
policy_builder (:py:class:`ModelBuilder[StochasticPolicy] <nnabla_rl.builders.ModelBuilder>`):
builder of actor models
policy_solver_builder (:py:class:`SolverBuilder <nnabla_rl.builders.SolverBuilder>`):
builder of policy solvers
replay_buffer_builder (:py:class:`ReplayBufferBuilder <nnabla_rl.builders.ReplayBufferBuilder>`):
builder of replay_buffer
explorer_builder (:py:class:`ExplorerBuilder <nnabla_rl.builders.ExplorerBuilder>`):
builder of environment explorer
"""
# type declarations to type check with mypy
# NOTE: declared variables are instance variable and NOT class variable, unless it is marked with ClassVar
# See https://mypy.readthedocs.io/en/stable/class_basics.html for details
_config: ICML2018SACConfig
_v: VFunction
_v_solver: nn.solver.Solver
_target_v: VFunction
_q1: QFunction
_q2: QFunction
_train_q_functions: List[QFunction]
_train_q_solvers: Dict[str, nn.solver.Solver]
_replay_buffer: ReplayBuffer
_explorer_builder: ExplorerBuilder
_environment_explorer: EnvironmentExplorer
_policy_trainer: ModelTrainer
_q_function_trainer: ModelTrainer
_v_function_trainer: ModelTrainer
_eval_state_var: nn.Variable
_eval_deterministic_action: nn.Variable
_eval_probabilistic_action: nn.Variable
_policy_trainer_state: Dict[str, Any]
_q_function_trainer_state: Dict[str, Any]
_v_function_trainer_state: Dict[str, Any]
def __init__(self, env_or_env_info: Union[gym.Env, EnvironmentInfo],
config: ICML2018SACConfig = ICML2018SACConfig(),
v_function_builder: ModelBuilder[VFunction] = DefaultVFunctionBuilder(),
v_solver_builder: SolverBuilder = DefaultSolverBuilder(),
q_function_builder: ModelBuilder[QFunction] = DefaultQFunctionBuilder(),
q_solver_builder: SolverBuilder = DefaultSolverBuilder(),
policy_builder: ModelBuilder[StochasticPolicy] = DefaultPolicyBuilder(),
policy_solver_builder: SolverBuilder = DefaultSolverBuilder(),
replay_buffer_builder: ReplayBufferBuilder = DefaultReplayBufferBuilder(),
explorer_builder: ExplorerBuilder = DefaultExplorerBuilder()):
super(ICML2018SAC, self).__init__(env_or_env_info, config=config)
self._explorer_builder = explorer_builder
with nn.context_scope(context.get_nnabla_context(self._config.gpu_id)):
self._v = v_function_builder(scope_name="v", env_info=self._env_info, algorithm_config=self._config)
self._v_solver = v_solver_builder(env_info=self._env_info, algorithm_config=self._config)
self._target_v = self._v.deepcopy('target_' + self._v.scope_name)
self._q1 = q_function_builder(scope_name="q1", env_info=self._env_info, algorithm_config=self._config)
self._q2 = q_function_builder(scope_name="q2", env_info=self._env_info, algorithm_config=self._config)
self._train_q_functions = [self._q1, self._q2]
self._train_q_solvers = {}
for q in self._train_q_functions:
self._train_q_solvers[q.scope_name] = q_solver_builder(
env_info=self._env_info, algorithm_config=self._config)
self._pi = policy_builder(scope_name="pi", env_info=self._env_info, algorithm_config=self._config)
self._pi_solver = policy_solver_builder(env_info=self._env_info, algorithm_config=self._config)
self._replay_buffer = replay_buffer_builder(env_info=self._env_info, algorithm_config=self._config)
self._evaluation_actor = _StochasticPolicyActionSelector(
self._env_info, self._pi.shallowcopy(), deterministic=True)
self._exploration_actor = _StochasticPolicyActionSelector(
self._env_info, self._pi.shallowcopy(), deterministic=False)
@eval_api
def compute_eval_action(self, state, *, begin_of_episode=False, extra_info={}):
with nn.context_scope(context.get_nnabla_context(self._config.gpu_id)):
action, _ = self._evaluation_action_selector(state, begin_of_episode=begin_of_episode)
return action
def _before_training_start(self, env_or_buffer):
# set context globally to ensure that the training runs on configured gpu
context.set_nnabla_context(self._config.gpu_id)
self._environment_explorer = self._setup_environment_explorer(env_or_buffer)
self._policy_trainer = self._setup_policy_training(env_or_buffer)
self._q_function_trainer = self._setup_q_function_training(env_or_buffer)
self._v_function_trainer = self._setup_v_function_training(env_or_buffer)
def _setup_environment_explorer(self, env_or_buffer):
return None if self._is_buffer(env_or_buffer) else self._explorer_builder(self._env_info, self._config, self)
def _setup_policy_training(self, env_or_buffer):
# NOTE: Fix temperature to 1.0. Because This version of SAC adjusts it by scaling the reward
policy_trainer_config = MT.policy_trainers.SoftPolicyTrainerConfig(
fixed_temperature=True,
unroll_steps=self._config.pi_unroll_steps,
burn_in_steps=self._config.pi_burn_in_steps,
reset_on_terminal=self._config.pi_reset_rnn_on_terminal)
temperature = MT.policy_trainers.soft_policy_trainer.AdjustableTemperature(
scope_name='temperature',
initial_value=1.0)
policy_trainer = MT.policy_trainers.SoftPolicyTrainer(
models=self._pi,
solvers={self._pi.scope_name: self._pi_solver},
q_functions=self._train_q_functions,
temperature=temperature,
temperature_solver=None,
env_info=self._env_info,
config=policy_trainer_config)
return policy_trainer
def _setup_q_function_training(self, env_or_buffer):
q_function_trainer_param = MT.q_value_trainers.VTargetedQTrainerConfig(
reduction_method='mean',
q_loss_scalar=0.5,
num_steps=self._config.num_steps,
unroll_steps=self._config.q_unroll_steps,
burn_in_steps=self._config.q_burn_in_steps,
reset_on_terminal=self._config.q_reset_rnn_on_terminal)
q_function_trainer = MT.q_value_trainers.VTargetedQTrainer(
train_functions=self._train_q_functions,
solvers=self._train_q_solvers,
target_functions=self._target_v,
env_info=self._env_info,
config=q_function_trainer_param)
return q_function_trainer
def _setup_v_function_training(self, env_or_buffer):
v_function_trainer_config = MT.v_value_trainers.SoftVTrainerConfig(
reduction_method='mean',
v_loss_scalar=0.5,
unroll_steps=self._config.v_unroll_steps,
burn_in_steps=self._config.v_burn_in_steps,
reset_on_terminal=self._config.v_reset_rnn_on_terminal)
v_function_trainer = MT.v_value_trainers.SoftVTrainer(
train_functions=self._v,
solvers={self._v.scope_name: self._v_solver},
target_functions=self._train_q_functions, # Set training q as target
target_policy=self._pi,
env_info=self._env_info,
config=v_function_trainer_config)
sync_model(self._v, self._target_v, 1.0)
return v_function_trainer
def _run_online_training_iteration(self, env):
for _ in range(self._config.environment_steps):
self._run_environment_step(env)
for _ in range(self._config.gradient_steps):
self._run_gradient_step(self._replay_buffer)
def _run_offline_training_iteration(self, buffer):
self._sac_training(buffer)
def _run_environment_step(self, env):
experiences = self._environment_explorer.step(env)
self._replay_buffer.append_all(experiences)
def _run_gradient_step(self, replay_buffer):
if self._config.start_timesteps < self.iteration_num:
self._sac_training(replay_buffer)
def _sac_training(self, replay_buffer):
pi_steps = self._config.pi_burn_in_steps + self._config.pi_unroll_steps
q_steps = self._config.num_steps + self._config.q_burn_in_steps + self._config.q_unroll_steps - 1
v_steps = self._config.v_burn_in_steps + self._config.v_unroll_steps
num_steps = max(pi_steps, max(q_steps, v_steps))
experiences_tuple, info = replay_buffer.sample(self._config.batch_size, num_steps=num_steps)
if num_steps == 1:
experiences_tuple = (experiences_tuple, )
assert len(experiences_tuple) == num_steps
batch = None
for experiences in reversed(experiences_tuple):
(s, a, r, non_terminal, s_next, rnn_states_dict, *_) = marshal_experiences(experiences)
rnn_states = rnn_states_dict['rnn_states'] if 'rnn_states' in rnn_states_dict else {}
batch = TrainingBatch(batch_size=self._config.batch_size,
s_current=s,
a_current=a,
gamma=self._config.gamma,
reward=r,
non_terminal=non_terminal,
s_next=s_next,
weight=info['weights'],
next_step_batch=batch,
rnn_states=rnn_states)
# Train in the order of v -> q -> policy
self._v_function_trainer_state = self._v_function_trainer.train(batch)
self._q_function_trainer_state = self._q_function_trainer.train(batch)
if self.iteration_num % self._config.target_update_interval == 0:
sync_model(self._v, self._target_v, tau=self._config.tau)
self._policy_trainer_state = self._policy_trainer.train(batch)
td_errors = self._q_function_trainer_state['td_errors']
replay_buffer.update_priorities(td_errors)
def _evaluation_action_selector(self, s, *, begin_of_episode=False):
return self._evaluation_actor(s, begin_of_episode=begin_of_episode)
def _exploration_action_selector(self, s, *, begin_of_episode=False):
return self._exploration_actor(s, begin_of_episode=begin_of_episode)
def _models(self):
models = [self._v, self._target_v, self._q1, self._q2, self._pi]
return {model.scope_name: model for model in models}
def _solvers(self):
solvers = {}
solvers.update(self._train_q_solvers)
solvers[self._v.scope_name] = self._v_solver
solvers[self._pi.scope_name] = self._pi_solver
return solvers
[docs] @classmethod
def is_rnn_supported(cls):
return True
[docs] @classmethod
def is_supported_env(cls, env_or_env_info):
env_info = EnvironmentInfo.from_env(env_or_env_info) if isinstance(env_or_env_info, gym.Env) \
else env_or_env_info
return not env_info.is_discrete_action_env() and not env_info.is_tuple_action_env()
@property
def latest_iteration_state(self):
latest_iteration_state = super(ICML2018SAC, self).latest_iteration_state
if hasattr(self, '_policy_trainer_state'):
latest_iteration_state['scalar'].update({'pi_loss': float(self._policy_trainer_state['pi_loss'])})
if hasattr(self, '_v_function_trainer_state'):
latest_iteration_state['scalar'].update({'v_loss': float(self._v_function_trainer_state['v_loss'])})
if hasattr(self, '_q_function_trainer_state'):
latest_iteration_state['scalar'].update({'q_loss': float(self._q_function_trainer_state['q_loss'])})
latest_iteration_state['histogram'].update(
{'td_errors': self._q_function_trainer_state['td_errors'].flatten()})
return latest_iteration_state
@property
def trainers(self):
return {
"v_function": self._v_function_trainer,
"q_function": self._q_function_trainer,
"policy": self._policy_trainer,
}