def create_d4rl_env_and_dataset(
task_name,
batch_size
):
"""Create gym environment and dataset for d4rl.
Args:
task_name: Name of d4rl task.
batch_size: Mini batch size.
Returns:
Gym env and dataset.
"""
env = gym.make(task_name)
env = wrappers.GymWrapper(env)
dataset = d4rl.qlearning_dataset(env)
states = np.array(dataset['observations'], dtype=np.float32)
actions = np.array(dataset['actions'], dtype=np.float32)
rewards = np.array(dataset['rewards'], dtype=np.float32)
discounts = np.array(np.logical_not(dataset['terminals']), dtype=np.float32)
next_states = np.array(dataset['next_observations'], dtype=np.float32)
dataset = tf_data.Dataset.from_tensor_slices(
Inputs(data=(states, actions, rewards, discounts, next_states))
).cache().shuffle(
states.shape[0], reshuffle_each_iteration=True
).repeat().batch(
batch_size, drop_remainder=True
).prefetch(tf_data.experimental.AUTOTUNE)
return env, dataset
def make_environment(suite: str, task: str) -> dm_env.Environment:
"""Makes the requested continuous control environment.
Args:
suite: One of 'gym' or 'control'.
task: Task to load. If `suite` is 'control', the task must be formatted as
f'{domain_name}:{task_name}'
Returns:
An environment satisfying the dm_env interface expected by Acme agents.
"""
if suite not in _VALID_TASK_SUITES:
raise ValueError(
f'Unsupported suite: {suite}. Expected one of {_VALID_TASK_SUITES}'
)
if suite == 'gym':
env = gym.make(task)
# Make sure the environment obeys the dm_env.Environment interface.
env = wrappers.GymWrapper(env)
elif suite == 'control':
# Load dm_suite lazily not require Mujoco license when not using it.
from dm_control import suite as dm_suite # pylint: disable=g-import-not-at-top
domain_name, task_name = task.split(':')
env = dm_suite.load(domain_name, task_name)
env = wrappers.ConcatObservationWrapper(env)
# Wrap the environment so the expected continuous action spec is [-1, 1].
# Note: this is a no-op on 'control' tasks.
env = wrappers.CanonicalSpecWrapper(env, clip=True)
env = wrappers.SinglePrecisionWrapper(env)
return env
def make_environment_atari(env_name, seed):
env = gym.make(env_name)
env = AtariWrapper(env,
partial_observation_wrapper=QuadrantObservationWrapper,
partial_percentage=1.0,
seed=seed)
env = wrappers.GymWrapper(env)
return env
def make_environment(
name: str,
seed: Optional[int] = None,
) -> dm_env.Environment:
env = gym.make(name)
env.seed(seed)
env = wrappers.GymWrapper(env)
return env
def make_gym_environment(
task_name: str = 'MountainCarContinuous-v0') -> dm_env.Environment:
"""Creates an OpenAI Gym environment."""
# Load the gym environment.
environment = gym.make(task_name)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
environment = wrappers.SinglePrecisionWrapper(environment)
return environment
def create_d4rl_env(
task_name,
):
"""Create the environment for the d4rl task.
Args:
task_name: Name of d4rl task.
Returns:
dm env.
"""
env = gym.make(task_name)
env = wrappers.GymWrapper(env)
return env
def make_environment(
task: str = 'MountainCarContinuous-v0') -> dm_env.Environment:
"""Creates an OpenAI Gym environment."""
# Load the gym environment.
environment = gym.make(task)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
# Clip the action returned by the agent to the environment spec.
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
environment = wrappers.SinglePrecisionWrapper(environment)
return environment
def make_environment(evaluation: bool = False,
task: str = 'HalfCheetah-v3') -> dm_env.Environment:
"""Creates an OpenAI Gym environment."""
del evaluation
# Load the gym environment.
environment = gym.make(task)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
# Clip the action returned by the agent to the environment spec.
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
environment = wrappers.SinglePrecisionWrapper(environment)
return environment
def make_single_agent_env(scenario: str, render=False):
scenario = SingleAgentScenario.from_spec(scenario, rendering=render)
env = VectorizedSingleAgentRaceEnv(scenarios=[scenario])
env = wrap_env(env=env, wrapper_configs='single_agent_wrappers.yml')
env = wrappers.GymWrapper(environment=env)
env = wrappers.SinglePrecisionWrapper(env)
return env
# def make_multi_agent_env(scenario: str, render=False, test=False):
# scenario = MultiAgentScenario.from_spec(scenario, rendering=render)
# env = VectorizedMultiAgentRaceEnv(scenarios=[scenario])
# if test:
# env = wrap_env(env=env, wrapper_configs='multi_agent_test_wrappers.yml')
# else:
# env = wrap_env(env=env, wrapper_configs='multi_agent_wrappers.yml')
#
# env = MultiAgentGymWrapper(environment=env)
# env = wrappers.SinglePrecisionWrapper(env)
# return env
def make_environment(task, evaluation = False):
"""Creates an OpenAI Gym environment."""
# Load the gym environment.
environment = gym.make(task)
environment = env_wrappers.AdroitSparseRewardWrapper(environment)
# Make sure the environment obeys the dm_env.Environment interface.
environment = wrappers.GymWrapper(environment)
# Clip the action returned by the agent to the environment spec.
environment = wrappers.CanonicalSpecWrapper(environment, clip=True)
environment = wrappers.SinglePrecisionWrapper(environment)
if evaluation:
environment = env_wrappers.SuccessRewardWrapper(environment,
success_threshold=1.)
return environment
from env.RSEnv import RSEnv from env.TestRSEnv import TestRSEnv from acme import environment_loop from acme import specs from acme import wrappers from acme.agents.tf import d4pg from acme.tf import networks from acme.tf import utils as tf2_utils from acme.utils import loggers import numpy as np import sonnet as snt import gym environment = RSEnv() environment = wrappers.GymWrapper(environment) # To dm_env interface. # Make sure the environment outputs single-precision floats. environment = wrappers.SinglePrecisionWrapper(environment) # Grab the spec of the environment. environment_spec = specs.make_environment_spec(environment) #@title Build agent networks # BUILDING A D4PG AGENT # Get total number of action dimensions from action spec. num_dimensions = np.prod(environment_spec.actions.shape, dtype=int) # Create the shared observation network; here simply a state-less operation. observation_network = tf2_utils.batch_concat
