Python DDPGAgentParameters.input_filter 예제들

예제 #1

파일: ControlSuite_DDPG.py 프로젝트: wwxFromTju/coach

    agent_params.network_wrappers['actor'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['critic'].input_embedders_parameters['measurements'] = \
    agent_params.network_wrappers['critic'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['actor'].input_embedders_parameters[
    'measurements'].scheme = [Dense(300)]
agent_params.network_wrappers['actor'].middleware_parameters.scheme = [
    Dense(200)
]
agent_params.network_wrappers['critic'].input_embedders_parameters[
    'measurements'].scheme = [Dense(400)]
agent_params.network_wrappers['critic'].middleware_parameters.scheme = [
    Dense(300)
]
agent_params.network_wrappers['critic'].input_embedders_parameters[
    'action'].scheme = EmbedderScheme.Empty
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter("rescale",
                                            RewardRescaleFilter(1 / 10.))

###############
# Environment #
###############
env_params = ControlSuiteEnvironmentParameters(
    level=SingleLevelSelection(control_suite_envs))

########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = ['cartpole:swingup', 'hopper:hop']

예제 #2

schedule_params.heatup_steps = EnvironmentSteps(1000)

#########
# Agent #
#########
agent_params = DDPGAgentParameters()
agent_params.network_wrappers['actor'].input_embedders_parameters['measurements'] = \
    agent_params.network_wrappers['actor'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['critic'].input_embedders_parameters['measurements'] = \
    agent_params.network_wrappers['critic'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['actor'].input_embedders_parameters['measurements'].scheme = [Dense([300])]
agent_params.network_wrappers['actor'].middleware_parameters.scheme = [Dense([200])]
agent_params.network_wrappers['critic'].input_embedders_parameters['measurements'].scheme = [Dense([400])]
agent_params.network_wrappers['critic'].middleware_parameters.scheme = [Dense([300])]
agent_params.network_wrappers['critic'].input_embedders_parameters['action'].scheme = EmbedderScheme.Empty
agent_params.input_filter = MujocoInputFilter()
agent_params.input_filter.add_reward_filter("rescale", RewardRescaleFilter(1/10.))

###############
# Environment #
###############
env_params = ControlSuiteEnvironmentParameters()
env_params.level = SingleLevelSelection(control_suite_envs)

vis_params = VisualizationParameters()
vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()]
vis_params.dump_mp4 = False

########
# Test #
########

예제 #3

파일: CARLA_3_Cameras_DDPG.py 프로젝트: bigdatasciencegroup/intel-ailab-reinforcement-learning-coach

agent_params.network_wrappers['critic'].input_embedders_parameters['forward_camera'] = \
    agent_params.network_wrappers['critic'].input_embedders_parameters.pop('observation')

# left camera
agent_params.network_wrappers['actor'].input_embedders_parameters['left_camera'] = \
    copy.deepcopy(agent_params.network_wrappers['actor'].input_embedders_parameters['forward_camera'])
agent_params.network_wrappers['critic'].input_embedders_parameters['left_camera'] = \
    copy.deepcopy(agent_params.network_wrappers['critic'].input_embedders_parameters['forward_camera'])

# right camera
agent_params.network_wrappers['actor'].input_embedders_parameters['right_camera'] = \
    copy.deepcopy(agent_params.network_wrappers['actor'].input_embedders_parameters['forward_camera'])
agent_params.network_wrappers['critic'].input_embedders_parameters['right_camera'] = \
    copy.deepcopy(agent_params.network_wrappers['critic'].input_embedders_parameters['forward_camera'])

agent_params.input_filter = CarlaInputFilter()
agent_params.input_filter.copy_filters_from_one_observation_to_another(
    'forward_camera', 'left_camera')
agent_params.input_filter.copy_filters_from_one_observation_to_another(
    'forward_camera', 'right_camera')

###############
# Environment #
###############
env_params = CarlaEnvironmentParameters()
env_params.cameras = [CameraTypes.FRONT, CameraTypes.LEFT, CameraTypes.RIGHT]

graph_manager = BasicRLGraphManager(agent_params=agent_params,
                                    env_params=env_params,
                                    schedule_params=schedule_params,
                                    vis_params=VisualizationParameters())