# 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, Union
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.algorithms.common_utils import _GreedyActionSelector
from nnabla_rl.builders import ExplorerBuilder, 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.model_trainers.model_trainer import ModelTrainer, TrainingBatch
from nnabla_rl.models import IQNQuantileFunction, StateActionQuantileFunction
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 IQNConfig(AlgorithmConfig):
"""List of configurations for IQN 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 0.00005.
batch_size (int): training batch size. Defaults to 32.
num_steps (int): number of steps for N-step Q targets. Defaults to 1.
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 (int): the interval of learner update. Defaults to 4.
target_update_frequency (int): 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.
N (int): Number of samples to compute the current state's quantile values. Defaults to 64.
N_prime (int): Number of samples to compute the target state's quantile values. Defaults to 64.
K (int): Number of samples to compute greedy next action. Defaults to 32.
kappa (float): threshold value of quantile huber loss. Defaults to 1.0.
embedding_dim (int): dimension of embedding for the sample point. Defaults to 64.
unroll_steps (int): Number of steps to unroll tranining network.
The network will be unrolled even though the provided model doesn't have RNN layers.
Defaults to 1.
burn_in_steps (int): Number of burn-in steps to initiaze recurrent layer states during training.
This flag does not take effect if given model is not an RNN model.
Defaults to 0.
reset_rnn_on_terminal (bool): Reset 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 True.
"""
gamma: float = 0.99
learning_rate: float = 0.00005
batch_size: int = 32
num_steps: int = 1
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
N: int = 64
N_prime: int = 64
K: int = 32
kappa: float = 1.0
embedding_dim: int = 64
# rnn model support
unroll_steps: int = 1
burn_in_steps: int = 0
reset_rnn_on_terminal: bool = True
def __post_init__(self):
"""__post_init__
Check that set values are in valid range.
"""
self._assert_between(self.gamma, 0.0, 1.0, 'gamma')
self._assert_positive(self.batch_size, 'batch_size')
self._assert_positive(self.replay_buffer_size, 'replay_buffer_size')
self._assert_positive(self.num_steps, 'num_steps')
self._assert_positive(self.learner_update_frequency, 'learner_update_frequency')
self._assert_positive(self.target_update_frequency, 'target_update_frequency')
self._assert_positive(self.max_explore_steps, 'max_explore_steps')
self._assert_positive(self.learning_rate, 'learning_rate')
self._assert_positive(self.initial_epsilon, 'initial_epsilon')
self._assert_positive(self.final_epsilon, 'final_epsilon')
self._assert_positive(self.test_epsilon, 'test_epsilon')
self._assert_positive(self.N, 'N')
self._assert_positive(self.N_prime, 'N_prime')
self._assert_positive(self.K, 'K')
self._assert_positive(self.kappa, 'kappa')
self._assert_positive(self.embedding_dim, 'embedding_dim')
self._assert_positive(self.unroll_steps, 'unroll_steps')
self._assert_positive_or_zero(self.burn_in_steps, 'burn_in_steps')
def risk_neutral_measure(tau):
return tau
class DefaultQuantileFunctionBuilder(ModelBuilder[StateActionQuantileFunction]):
def build_model(self, # type: ignore[override]
scope_name: str,
env_info: EnvironmentInfo,
algorithm_config: IQNConfig,
**kwargs) -> StateActionQuantileFunction:
assert isinstance(algorithm_config, IQNConfig)
risk_measure_function = kwargs['risk_measure_function']
return IQNQuantileFunction(scope_name,
env_info.action_dim,
algorithm_config.embedding_dim,
K=algorithm_config.K,
risk_measure_function=risk_measure_function)
class DefaultSolverBuilder(SolverBuilder):
def build_solver(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: IQNConfig,
**kwargs) -> nn.solver.Solver:
assert isinstance(algorithm_config, IQNConfig)
return NS.Adam(alpha=algorithm_config.learning_rate, eps=1e-2 / algorithm_config.batch_size)
class DefaultReplayBufferBuilder(ReplayBufferBuilder):
def build_replay_buffer(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: IQNConfig,
**kwargs) -> ReplayBuffer:
assert isinstance(algorithm_config, IQNConfig)
return ReplayBuffer(capacity=algorithm_config.replay_buffer_size)
class DefaultExplorerBuilder(ExplorerBuilder):
def build_explorer(self, # type: ignore[override]
env_info: EnvironmentInfo,
algorithm_config: IQNConfig,
algorithm: "IQN",
**kwargs) -> EnvironmentExplorer:
explorer_config = EE.LinearDecayEpsilonGreedyExplorerConfig(
warmup_random_steps=algorithm_config.start_timesteps,
initial_step_num=algorithm.iteration_num,
initial_epsilon=algorithm_config.initial_epsilon,
final_epsilon=algorithm_config.final_epsilon,
max_explore_steps=algorithm_config.max_explore_steps
)
explorer = EE.LinearDecayEpsilonGreedyExplorer(
greedy_action_selector=algorithm._exploration_action_selector,
random_action_selector=algorithm._random_action_selector,
env_info=env_info,
config=explorer_config)
return explorer
[docs]class IQN(Algorithm):
"""Implicit Quantile Network algorithm.
This class implements the Implicit Quantile Network (IQN) algorithm
proposed by W. Dabney, et al. in the paper: "Implicit Quantile Networks for Distributional Reinforcement Learning"
For details see: https://arxiv.org/pdf/1806.06923.pdf
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:`IQNConfig <nnabla_rl.algorithms.iqn.IQNConfig>`): configuration of IQN algorithm
quantile_function_builder (:py:class:`ModelBuilder[StateActionQuantileFunction] \
<nnabla_rl.builders.ModelBuilder>`): buider of state-action quantile function models
quantile_solver_builder (:py:class:`SolverBuilder <nnabla_rl.builders.SolverBuilder>`):
builder for state action quantile function 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: IQNConfig
_quantile_function: StateActionQuantileFunction
_target_quantile_function: StateActionQuantileFunction
_quantile_function_solver: nn.solver.Solver
_replay_buffer: ReplayBuffer
_explorer_builder: ExplorerBuilder
_environment_explorer: EnvironmentExplorer
_quantile_function_trainer: ModelTrainer
_quantile_function_trainer_state: Dict[str, Any]
def __init__(self, env_or_env_info: Union[gym.Env, EnvironmentInfo],
config: IQNConfig = IQNConfig(),
quantile_function_builder: ModelBuilder[StateActionQuantileFunction]
= DefaultQuantileFunctionBuilder(),
quantile_solver_builder: SolverBuilder = DefaultSolverBuilder(),
replay_buffer_builder: ReplayBufferBuilder = DefaultReplayBufferBuilder(),
explorer_builder: ExplorerBuilder = DefaultExplorerBuilder(),
risk_measure_function=risk_neutral_measure):
super(IQN, 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)):
kwargs = {}
kwargs['risk_measure_function'] = risk_measure_function
self._quantile_function = quantile_function_builder(
'quantile_function', self._env_info, self._config, **kwargs)
self._target_quantile_function = self._quantile_function.deepcopy('target_quantile_function')
self._quantile_function_solver = quantile_solver_builder(self._env_info, self._config)
self._replay_buffer = replay_buffer_builder(self._env_info, self._config)
self._evaluation_actor = _GreedyActionSelector(
self._env_info, self._quantile_function.shallowcopy().as_q_function())
self._exploration_actor = _GreedyActionSelector(
self._env_info, self._quantile_function.shallowcopy().as_q_function())
@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, _), _ = epsilon_greedy_action_selection(state,
self._evaluation_action_selector,
self._random_action_selector,
epsilon=self._config.test_epsilon,
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._quantile_function_trainer = self._setup_quantile_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_quantile_function_training(self, env_or_buffer):
trainer_config = MT.q_value_trainers.IQNQTrainerConfig(
num_steps=self._config.num_steps,
N=self._config.N,
N_prime=self._config.N_prime,
K=self._config.K,
kappa=self._config.kappa,
unroll_steps=self._config.unroll_steps,
burn_in_steps=self._config.burn_in_steps,
reset_on_terminal=self._config.reset_rnn_on_terminal)
quantile_function_trainer = MT.q_value_trainers.IQNQTrainer(
train_functions=self._quantile_function,
solvers={self._quantile_function.scope_name: self._quantile_function_solver},
target_function=self._target_quantile_function,
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._quantile_function, self._target_quantile_function)
return quantile_function_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._iqn_training(self._replay_buffer)
def _run_offline_training_iteration(self, buffer):
self._iqn_training(buffer)
def _iqn_training(self, replay_buffer):
num_steps = self._config.num_steps + self._config.burn_in_steps + self._config.unroll_steps - 1
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)
self._quantile_function_trainer_state = self._quantile_function_trainer.train(batch)
if self.iteration_num % self._config.target_update_frequency:
sync_model(self._quantile_function, self._target_quantile_function)
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 _random_action_selector(self, s, *, begin_of_episode=False):
action = self._env_info.action_space.sample()
return np.asarray(action).reshape((1, )), {}
def _models(self):
models = {}
models[self._quantile_function.scope_name] = self._quantile_function
return models
def _solvers(self):
solvers = {}
solvers[self._quantile_function.scope_name] = self._quantile_function_solver
return solvers
[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_continuous_action_env() and not env_info.is_tuple_action_env()
[docs] @classmethod
def is_rnn_supported(self):
return True
@property
def latest_iteration_state(self):
latest_iteration_state = super(IQN, self).latest_iteration_state
if hasattr(self, '_quantile_function_trainer_state'):
latest_iteration_state['scalar'].update({'q_loss': float(self._quantile_function_trainer_state['q_loss'])})
return latest_iteration_state
@property
def trainers(self):
return {"q_function": self._quantile_function_trainer}