def __init__(self,
state_normalization_parameters: Dict[str,
NormalizationParameters],
parameters: Union[DiscreteActionModelParameters,
ContinuousActionModelParameters],
skip_normalization: Optional[bool] = False) -> None:
print(state_normalization_parameters)
print(parameters)
self._state_normalization_parameters = state_normalization_parameters
MLTrainer.__init__(self, "rl_trainer", parameters.training)
self.target_network = TargetNetwork(self,
parameters.rl.target_update_rate)
self.reward_burnin = parameters.rl.reward_burnin
self.maxq_learning = parameters.rl.maxq_learning
self.rl_discount_rate = parameters.rl.gamma
self.training_iteration = 0
self._buffers = None
self.minibatch_size = parameters.training.minibatch_size
self.skip_normalization = skip_normalization
self._prepare_state_normalization()
def __init__(self, name, parameters, scaled_output=True):
"""
:param name: A unique name for this trainer used to create the data on the
caffe2 workspace
:param layers: A list of integers describing the layer sizes
:param activations: A list of strings describing the activation functions
"""
self.scaled_output = scaled_output
MLTrainer.__init__(self, name, parameters)
def __init__(self, name: str, fc_parameters: TrainingParameters,
cnn_parameters: CNNModelParameters, img_height: int,
img_width: int) -> None:
self.init_height = img_height
self.init_width = img_width
self.dims = cnn_parameters.conv_dims
self.conv_height_kernels = cnn_parameters.conv_height_kernels
self.conv_width_kernels = cnn_parameters.conv_width_kernels
self.pool_kernels_strides = cnn_parameters.pool_kernels_strides
self.pool_types = cnn_parameters.pool_types
MLTrainer.__init__(self, name, fc_parameters)
def __init__(
self,
parameters: Union[DiscreteActionModelParameters,
ContinuousActionModelParameters],
) -> None:
logger.info(str(parameters))
assert parameters.training.layers[0] >= 0,\
"Set layers[0] to a the number of features"
self.num_features = parameters.training.layers[0]
MLTrainer.__init__(self, RL_TRAINER_MODEL_ID, parameters.training)
self.target_network = TargetNetwork(
self, parameters.rl.target_update_rate
)
self.reward_burnin = parameters.rl.reward_burnin
self.maxq_learning = parameters.rl.maxq_learning
self.rl_discount_rate = parameters.rl.gamma
self.rl_temperature = parameters.rl.temperature
self.training_iteration = 0
self.minibatch_size = parameters.training.minibatch_size
self.parameters = parameters
self.loss_blob: Optional[str] = None
workspace.FeedBlob('states', np.array([0], dtype=np.float32))
workspace.FeedBlob('actions', np.array([0], dtype=np.float32))
workspace.FeedBlob('rewards', np.array([0], dtype=np.float32))
workspace.FeedBlob('next_states', np.array([0], dtype=np.float32))
workspace.FeedBlob('not_terminals', np.array([0], dtype=np.float32))
workspace.FeedBlob('next_actions', np.array([0], dtype=np.float32))
workspace.FeedBlob(
'possible_next_actions', np.array([0], dtype=np.float32)
)
workspace.FeedBlob(
'possible_next_actions_lengths', np.array([0], dtype=np.float32)
)
self.rl_train_model: Optional[ModelHelper] = None
self.reward_train_model: Optional[ModelHelper] = None
self.q_score_model: Optional[ModelHelper] = None
self._create_reward_train_net()
self._create_rl_train_net()
self._create_q_score_net()
assert self.rl_train_model is not None
assert self.reward_train_model is not None
assert self.q_score_model is not None
def __init__(
self,
parameters: Union[DiscreteActionModelParameters,
ContinuousActionModelParameters],
) -> None:
logger.info(str(parameters))
assert parameters.training.layers[0] >= 0,\
"Set layers[0] to a the number of features"
self.num_features = parameters.training.layers[0]
MLTrainer.__init__(self, "rl_trainer", parameters.training)
self.target_network = TargetNetwork(self,
parameters.rl.target_update_rate)
self.reward_burnin = parameters.rl.reward_burnin
self.maxq_learning = parameters.rl.maxq_learning
self.rl_discount_rate = parameters.rl.gamma
self.training_iteration = 0
self.minibatch_size = parameters.training.minibatch_size