# Copyright 2020,2021 Sony Corporation.
# Copyright 2021 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, Union, cast
import gym
import numpy as np
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.builders import ModelBuilder, ReplayBufferBuilder, SolverBuilder
from nnabla_rl.environment_explorer import EnvironmentExplorer
from nnabla_rl.environment_explorers.epsilon_greedy_explorer import epsilon_greedy_action_selection
from nnabla_rl.environments.environment_info import EnvironmentInfo
from nnabla_rl.exceptions import UnsupportedEnvironmentException
from nnabla_rl.model_trainers.model_trainer import ModelTrainer, TrainingBatch
from nnabla_rl.models import C51ValueDistributionFunction, ValueDistributionFunction
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 CategoricalDQNConfig(AlgorithmConfig):
'''CategoricalDQNConfig
List of configurations for CategoricalDQN 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 0.001.
batch_size (int): training atch size. Defaults to 32.
start_timesteps (int): the timestep when training starts.\
The algorithm will collect experiences from the environment by acting randomly until this timestep.
Defaults to 50000.
replay_buffer_size (int): the capacity of replay buffer. Defaults to 1000000.
learner_update_frequency (float): the interval of learner update. Defaults to 4
target_update_frequency (float): the interval of target q-function update. Defaults to 10000.
max_explore_steps (int): the number of steps decaying the epsilon value.\
The epsilon will be decayed linearly \
:math:`\\epsilon=\\epsilon_{init} - step\\times\\frac{\\epsilon_{init} - \
\\epsilon_{final}}{max\\_explore\\_steps}`.\
Defaults to 1000000.
initial_epsilon (float): the initial epsilon value for ε-greedy explorer. Defaults to 1.0.
final_epsilon (float): the last epsilon value for ε-greedy explorer. Defaults to 0.01.
test_epsilon (float): the epsilon value on testing. Defaults to 0.001.
v_min (float): lower limit of the value used in value distribution function. Defaults to -10.0.
v_max (float): upper limit of the value used in value distribution function. Defaults to 10.0.
num_atoms (int): the number of bins used in value distribution function. Defaults to 51.
'''
gamma: float = 0.99
learning_rate: float = 0.00025
batch_size: int = 32
start_timesteps: int = 50000
replay_buffer_size: int = 1000000
learner_update_frequency: int = 4
target_update_frequency: int = 10000
max_explore_steps: int = 1000000
initial_epsilon: float = 1.0
final_epsilon: float = 0.01
test_epsilon: float = 0.001
v_min: float = -10.0
v_max: float = 10.0
num_atoms: int = 51
class DefaultValueDistFunctionBuilder(ModelBuilder[ValueDistributionFunction]):
def build_model(self, # type: ignore[override]
scope_name: str,
env_info: EnvironmentInfo,
algorithm_config: CategoricalDQNConfig,
**kwargs) -> ValueDistributionFunction:
return C51ValueDistributionFunction(scope_name,
env_info.action_dim,
algorithm_config.num_atoms,
algorithm_config.v_min,
algorithm_config.v_max)
class DefaultReplayBufferBuilder(ReplayBufferBuilder):
def build_replay_buffer(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: CategoricalDQNConfig,
**kwargs) -> ReplayBuffer:
return ReplayBuffer(capacity=algorithm_config.replay_buffer_size)
class DefaultSolverBuilder(SolverBuilder):
def build_solver(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: CategoricalDQNConfig,
**kwargs) -> nn.solver.Solver:
return NS.Adam(alpha=algorithm_config.learning_rate, eps=1e-2 / algorithm_config.batch_size)
[docs]class CategoricalDQN(Algorithm):
'''Categorical DQN algorithm.
This class implements the Categorical DQN algorithm
proposed by M. Bellemare, et al. in the paper: "A Distributional Perspective on Reinfocement Learning"
For details see: https://arxiv.org/abs/1707.06887
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:`CategoricalDQNConfig <nnabla_rl.algorithms.categorical_dqn.CategoricalDQNConfig>`):
configuration of the CategoricalDQN algorithm
value_distribution_builder (:py:class:`ModelBuilder[ValueDistributionFunctionFunction] \
<nnabla_rl.builders.ModelBuilder>`): builder of value distribution function models
value_distribution_solver_builder (:py:class:`SolverBuilder <nnabla_rl.builders.SolverBuilder>`):
builder of value distribution function solvers
replay_buffer_builder (:py:class:`ReplayBufferBuilder <nnabla_rl.builders.ReplayBufferBuilder>`):
builder of replay_buffer
'''
# 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: CategoricalDQNConfig
_atom_p: ValueDistributionFunction
_atom_p_solver: nn.solver.Solver
_target_atom_p: ValueDistributionFunction
_replay_buffer: ReplayBuffer
_environment_explorer: EnvironmentExplorer
_model_trainer: ModelTrainer
_eval_state_var: nn.Variable
_a_greedy: nn.Variable
_model_trainer_state: Dict[str, Any]
def __init__(self, env_or_env_info: Union[gym.Env, EnvironmentInfo],
config: CategoricalDQNConfig = CategoricalDQNConfig(),
value_distribution_builder: ModelBuilder[ValueDistributionFunction]
= DefaultValueDistFunctionBuilder(),
value_distribution_solver_builder: SolverBuilder = DefaultSolverBuilder(),
replay_buffer_builder: ReplayBufferBuilder = DefaultReplayBufferBuilder()):
super(CategoricalDQN, self).__init__(env_or_env_info, config=config)
if not self._env_info.is_discrete_action_env():
raise UnsupportedEnvironmentException('{} only supports discrete action environment'.format(self.__name__))
with nn.context_scope(context.get_nnabla_context(self._config.gpu_id)):
self._atom_p = value_distribution_builder('atom_p_train', self._env_info, self._config)
self._atom_p_solver = value_distribution_solver_builder(self._env_info, self._config)
self._target_atom_p = cast(ValueDistributionFunction, self._atom_p.deepcopy('target_atom_p_train'))
self._replay_buffer = replay_buffer_builder(self._env_info, self._config)
@eval_api
def compute_eval_action(self, state):
with nn.context_scope(context.get_nnabla_context(self._config.gpu_id)):
(action, _), _ = epsilon_greedy_action_selection(state,
self._greedy_action_selector,
self._random_action_selector,
epsilon=self._config.test_epsilon)
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._model_trainer = self._setup_value_distribution_function_training(env_or_buffer)
def _setup_environment_explorer(self, env_or_buffer):
if self._is_buffer(env_or_buffer):
return None
explorer_config = EE.LinearDecayEpsilonGreedyExplorerConfig(
warmup_random_steps=self._config.start_timesteps,
initial_step_num=self.iteration_num,
initial_epsilon=self._config.initial_epsilon,
final_epsilon=self._config.final_epsilon,
max_explore_steps=self._config.max_explore_steps
)
explorer = EE.LinearDecayEpsilonGreedyExplorer(
greedy_action_selector=self._greedy_action_selector,
random_action_selector=self._random_action_selector,
env_info=self._env_info,
config=explorer_config)
return explorer
def _setup_value_distribution_function_training(self, env_or_buffer):
trainer_config = MT.q_value_trainers.CategoricalDQNQTrainerConfig(
v_min=self._config.v_min,
v_max=self._config.v_max,
num_atoms=self._config.num_atoms)
model_trainer = MT.q_value_trainers.CategoricalDQNQTrainer(
train_functions=self._atom_p,
solvers={self._atom_p.scope_name: self._atom_p_solver},
target_function=self._target_atom_p,
env_info=self._env_info,
config=trainer_config)
# NOTE: Copy initial parameters after setting up the training
# Because the parameter is created after training graph construction
sync_model(self._atom_p, self._target_atom_p)
return model_trainer
def _run_online_training_iteration(self, env):
experiences = self._environment_explorer.step(env)
self._replay_buffer.append_all(experiences)
if self._config.start_timesteps < self.iteration_num:
if self.iteration_num % self._config.learner_update_frequency == 0:
self._categorical_dqn_training(self._replay_buffer)
def _run_offline_training_iteration(self, buffer):
self._categorical_dqn_training(buffer)
def _categorical_dqn_training(self, replay_buffer):
experiences, info = replay_buffer.sample(self._config.batch_size)
(s, a, r, non_terminal, s_next, *_) = marshal_experiences(experiences)
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'])
self._model_trainer_state = self._model_trainer.train(batch)
if self.iteration_num % self._config.target_update_frequency == 0:
sync_model(self._atom_p, self._target_atom_p)
td_errors = np.abs(self._model_trainer_state['td_errors'])
replay_buffer.update_priorities(td_errors)
@eval_api
def _greedy_action_selector(self, s):
s = np.expand_dims(s, axis=0)
if not hasattr(self, '_eval_state_var'):
self._eval_state_var = nn.Variable(s.shape)
q_function = self._atom_p.as_q_function()
self._a_greedy = q_function.argmax_q(self._eval_state_var)
self._eval_state_var.d = s
self._a_greedy.forward()
return np.squeeze(self._a_greedy.d, axis=0), {}
def _random_action_selector(self, s):
action = self._env_info.action_space.sample()
return np.asarray(action).reshape((1, )), {}
def _models(self):
models = {}
models[self._atom_p.scope_name] = self._atom_p
return models
def _solvers(self):
solvers = {}
solvers[self._atom_p.scope_name] = self._atom_p_solver
return solvers
@property
def latest_iteration_state(self):
latest_iteration_state = super(CategoricalDQN, self).latest_iteration_state
if hasattr(self, '_model_trainer_state'):
latest_iteration_state['scalar'].update(
{'cross_entropy_loss': self._model_trainer_state['cross_entropy_loss']})
latest_iteration_state['histogram'].update({'td_errors': self._model_trainer_state['td_errors'].flatten()})
return latest_iteration_state