def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='leaky_relu'),
'measurements': InputEmbedderParameters(activation_function='leaky_relu'),
'goal': InputEmbedderParameters(activation_function='leaky_relu')}
self.input_embedders_parameters['observation'].scheme = [
Conv2d(32, 8, 4),
Conv2d(64, 4, 2),
Conv2d(64, 3, 1),
Dense(512),
]
self.input_embedders_parameters['measurements'].scheme = [
Dense(128),
Dense(128),
Dense(128),
]
self.input_embedders_parameters['goal'].scheme = [
Dense(128),
Dense(128),
Dense(128),
]
self.middleware_parameters = FCMiddlewareParameters(activation_function='leaky_relu',
scheme=MiddlewareScheme.Empty)
self.heads_parameters = [MeasurementsPredictionHeadParameters(activation_function='leaky_relu')]
self.async_training = False
self.batch_size = 64
self.adam_optimizer_beta1 = 0.95
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [DNDQHeadParameters()]
self.optimizer_type = 'Adam'
self.should_get_softmax_probabilities = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [VHeadParameters(loss_weight=0.5), PolicyHeadParameters(loss_weight=1.0)]
self.optimizer_type = 'Adam'
self.clip_gradients = 40.0
self.async_training = True
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters()
}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [DNDQHeadParameters()]
self.optimizer_type = 'Adam'
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters()
}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [PolicyHeadParameters()]
self.async_training = True
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters(scheme=MiddlewareScheme.Medium)
self.heads_parameters = [RegressionHeadParameters()]
self.optimizer_type = 'Adam'
self.batch_size = 32
self.replace_mse_with_huber_loss = False
self.create_target_network = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [QHeadParameters()]
self.optimizer_type = 'Adam'
self.async_training = True
self.shared_optimizer = True
self.create_target_network = True
def test_fc_middleware():
params = FCMiddlewareParameters(scheme=MiddlewareScheme.Medium)
mid = FCMiddleware(params=params)
mid.initialize()
embedded_data = mx.nd.random.uniform(low=0, high=1, shape=(10, 100))
output = mid(embedded_data)
assert output.ndim == 2 # since last block was flatten
assert output.shape[0] == 10 # since batch_size is 10
assert output.shape[
1] == 512 # since last layer of middleware (middle scheme) had 512 units
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='leaky_relu',
input_rescaling={'image': 1.0})}
self.middleware_parameters = FCMiddlewareParameters(scheme=MiddlewareScheme.Empty)
self.heads_parameters = [RNDHeadParameters()]
self.create_target_network = False
self.optimizer_type = 'Adam'
self.batch_size = 100
self.learning_rate = 0.0001
self.should_get_softmax_probabilities = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters()
}
self.middleware_parameters = FCMiddlewareParameters(
activation_function='none')
self.heads_parameters = [RNDHeadParameters()]
self.optimizer_type = 'Adam'
self.clip_gradients = None
self.create_target_network = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(scheme=EmbedderScheme.Empty)}
self.middleware_parameters = FCMiddlewareParameters(scheme=MiddlewareScheme.Empty)
self.heads_parameters = [SACQHeadParameters()] # SACQHeadParameters includes the topology of the head
self.rescale_gradient_from_head_by_factor = [1]
self.optimizer_type = 'Adam'
self.batch_size = 256
self.async_training = False
self.learning_rate = 0.0003
self.create_target_network = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='relu')}
self.middleware_parameters = FCMiddlewareParameters(activation_function='relu')
self.heads_parameters = [SACPolicyHeadParameters()]
self.rescale_gradient_from_head_by_factor = [1]
self.optimizer_type = 'Adam'
self.batch_size = 256
self.async_training = False
self.learning_rate = 0.0003
self.create_target_network = False
self.l2_regularization = 0 # weight decay regularization. not used in the original paper
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='relu')}
self.middleware_parameters = FCMiddlewareParameters(activation_function='relu')
self.heads_parameters = [VHeadParameters(initializer='xavier')]
self.rescale_gradient_from_head_by_factor = [1]
self.optimizer_type = 'Adam'
self.batch_size = 256
self.async_training = False
self.learning_rate = 0.0003 # 3e-4 see appendix D in the paper
# tau is set in SoftActorCriticAlgorithmParameters.rate_for_copying_weights_to_target
self.create_target_network = True
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters(activation_function='tanh')
}
self.middleware_parameters = FCMiddlewareParameters(
activation_function='tanh')
self.heads_parameters = [VHeadParameters()]
self.async_training = True
self.l2_regularization = 0
self.create_target_network = True
self.batch_size = 128
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters(batchnorm=True)
}
self.middleware_parameters = FCMiddlewareParameters(batchnorm=True)
self.heads_parameters = [DDPGActorHeadParameters()]
self.optimizer_type = 'Adam'
self.batch_size = 64
self.async_training = False
self.learning_rate = 0.0001
self.create_target_network = True
self.shared_optimizer = True
self.scale_down_gradients_by_number_of_workers_for_sync_training = False
def __init__(self, num_q_networks):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(),
'action': InputEmbedderParameters(scheme=EmbedderScheme.Shallow)}
self.middleware_parameters = FCMiddlewareParameters(num_streams=num_q_networks)
self.heads_parameters = [TD3VHeadParameters()]
self.optimizer_type = 'Adam'
self.adam_optimizer_beta2 = 0.999
self.optimizer_epsilon = 1e-8
self.batch_size = 100
self.async_training = False
self.learning_rate = 0.001
self.create_target_network = True
self.shared_optimizer = True
self.scale_down_gradients_by_number_of_workers_for_sync_training = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='tanh')}
self.middleware_parameters = FCMiddlewareParameters(activation_function='tanh')
self.heads_parameters = [VHeadParameters(), PPOHeadParameters()]
self.batch_size = 64
self.optimizer_type = 'Adam'
self.clip_gradients = None
self.use_separate_networks_per_head = True
self.async_training = False
self.l2_regularization = 0
# The target network is used in order to freeze the old policy, while making updates to the new one
# in train_network()
self.create_target_network = True
self.shared_optimizer = True
self.scale_down_gradients_by_number_of_workers_for_sync_training = True
def create_middle_embedder(scheme_dict, embedder_type, activation_function):
"""Creates rl coach middleware
scheme_dict - Dictionary containing all the information required by
the scheme creation methods defined above.
embedder_type - String indicating desired inputembedder type, available
types are defined in SCHEME_TYPE
activation_function - Desired activationfunction for the embdedder
"""
try:
if not ActivationFunctions.has_activation_function(activation_function):
raise Exception("Invalid activation function for middleware")
scheme = SCHEME_TYPE[embedder_type](scheme_dict)
return FCMiddlewareParameters(scheme=scheme, activation_function=activation_function)
except KeyError as err:
raise Exception("Middleware, key {} not found".format(err.args[0]))
except Exception as err:
raise Exception("Error while creating middleware: {}".format(err))
# Agent Params #
################
agent_params = DDPGAgentParameters()
# actor
actor_network = agent_params.network_wrappers['actor']
actor_network.learning_rate = 0.001
actor_network.batch_size = 256
actor_network.optimizer_epsilon = 1e-08
actor_network.adam_optimizer_beta1 = 0.9
actor_network.adam_optimizer_beta2 = 0.999
actor_network.input_embedders_parameters = {
'observation': InputEmbedderParameters(scheme=EmbedderScheme.Empty),
'desired_goal': InputEmbedderParameters(scheme=EmbedderScheme.Empty)
}
actor_network.middleware_parameters = FCMiddlewareParameters(scheme=[Dense(256), Dense(256), Dense(256)])
actor_network.heads_parameters[0].batchnorm = False
# critic
critic_network = agent_params.network_wrappers['critic']
critic_network.learning_rate = 0.001
critic_network.batch_size = 256
critic_network.optimizer_epsilon = 1e-08
critic_network.adam_optimizer_beta1 = 0.9
critic_network.adam_optimizer_beta2 = 0.999
critic_network.input_embedders_parameters = {
'action': InputEmbedderParameters(scheme=EmbedderScheme.Empty),
'desired_goal': InputEmbedderParameters(scheme=EmbedderScheme.Empty),
'observation': InputEmbedderParameters(scheme=EmbedderScheme.Empty)
}
critic_network.middleware_parameters = FCMiddlewareParameters(scheme=[Dense(256), Dense(256), Dense(256)])
def __init__(self):
super().__init__()
self.heads_parameters = [RainbowQHeadParameters()]
self.middleware_parameters = FCMiddlewareParameters(
scheme=MiddlewareScheme.Empty)
activation_function='none'),
'lidar': InputEmbedderParameters(
scheme=[
Dense(256),
BatchnormActivationDropout(activation_function='relu', dropout_rate=0.5),
Dense(256),
BatchnormActivationDropout(activation_function='relu', dropout_rate=0.5)
],
activation_function='none') # we define the activation function for each layer explicitly
}
agent_params.network_wrappers['main'].middleware_parameters = \
FCMiddlewareParameters(
scheme=[
Dense(256),
BatchnormActivationDropout(activation_function='relu', dropout_rate=0.5),
Dense(128),
BatchnormActivationDropout(activation_function='relu', dropout_rate=0.5)
],
activation_function='none'
)
agent_params.network_wrappers['main'].learning_rate = 0.0003
# agent_params.network_wrappers['main'].input_embedders_parameters['front_camera'].activation_function = 'relu'
# agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'relu'
agent_params.network_wrappers['main'].batch_size = 128
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
agent_params.network_wrappers['main'].learning_rate_decay_steps = 50000
agent_params.network_wrappers['main'].learning_rate_decay_rate = 0.95
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 1000000)
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(10000000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(100)
schedule_params.evaluation_steps = EnvironmentEpisodes(3)
schedule_params.heatup_steps = EnvironmentSteps(0)
#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.beta_entropy = 0.05
agent_params.network_wrappers['main'].middleware_parameters = FCMiddlewareParameters()
agent_params.network_wrappers['main'].learning_rate = 0.0001
###############
# Environment #
###############
env_params = Atari(level=SingleLevelSelection(atari_deterministic_v4))
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = ['breakout', 'pong', 'space_invaders']
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=VisualizationParameters(),
InputEmbedderParameters(
scheme=[Conv2d(32, 8, 4),
Conv2d(32, 4, 2),
Conv2d(64, 4, 2)],
activation_function='relu',
dropout_rate=0.3),
'LIDAR':
InputEmbedderParameters(scheme=[Dense(64), Dense(32)],
activation_function='relu',
dropout_rate=0.3)
}
agent_params.network_wrappers['main'].middleware_parameters = \
FCMiddlewareParameters(
scheme=[
Dense(256)
],
activation_function='relu', dropout_rate=0.3
)
agent_params.network_wrappers['main'].learning_rate = 0.0003
#agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'relu'
agent_params.network_wrappers['main'].batch_size = 64
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
# agent_params.network_wrappers['main'].learning_rate_decay_steps = 60000
# agent_params.network_wrappers['main'].learning_rate_decay_rate = 0.95
# agent_params.network_wrappers['main'].input_embedders_parameters['observation'].batchnorm = True
# agent_params.network_wrappers['main'].input_embedders_parameters['observation'].dropout_rate = 0.3
# agent_params.network_wrappers['main'].l2_regularization = 2e-5
agent_params.algorithm.beta_entropy = 0.001
Conv2d(64, 3, 1),
Conv2d(64, 3, 2),
Conv2d(128, 3, 1)
])
agent_params.network_wrappers['main'].input_embedders_parameters = {
'observation': input_embedder_params
}
agent_params.network_wrappers['constant'].input_embedders_parameters = {
'observation': input_embedder_params
}
agent_params.network_wrappers['predictor'].input_embedders_parameters = {
'observation': input_embedder_params
}
middleware_scheme = MiddlewareScheme.Medium
agent_params.network_wrappers[
'main'].middleware_parameters = FCMiddlewareParameters(
scheme=middleware_scheme)
agent_params.network_wrappers[
'constant'].middleware_parameters = FCMiddlewareParameters(
activation_function='none', scheme=middleware_scheme)
agent_params.network_wrappers[
'predictor'].middleware_parameters = FCMiddlewareParameters(
activation_function='none', scheme=middleware_scheme)
agent_params.network_wrappers['main'].heads_parameters = [
DuelingQHeadParameters()
]
# ER
agent_params.memory.max_size = (MemoryGranularity.Transitions, 4096)
################
# Environment #
dropout_rate=0.5),
Dense(128),
BatchnormActivationDropout(activation_function='tanh',
dropout_rate=0.5)
],
activation_function=
'none' # we define the activation function for each layer explicitly
)
}
# simple fc middleware
agent_params.network_wrappers['main'].middleware_parameters = \
FCMiddlewareParameters(
scheme=[
Dense(512),
BatchnormActivationDropout(activation_function='tanh', dropout_rate=0.5)
],
activation_function='none'
)
# output branches
agent_params.network_wrappers['main'].heads_parameters = [
RegressionHeadParameters(
scheme=[
Dense(256),
BatchnormActivationDropout(activation_function='tanh',
dropout_rate=0.5),
Dense(256),
BatchnormActivationDropout(activation_function='tanh')
],
num_output_head_copies=4 # follow lane, left, right, straight
