def __init__()

in Experiments/PolicyManagers.py [0:0]

• 70 lines of code
• 5 McCabe index (conditional complexity)

	def __init__(self, number_policies=4, dataset=None, args=None):

		super(PolicyManager_Joint, self).__init__()

		self.args = args
		self.data = self.args.data
		self.number_policies = number_policies
		self.latent_z_dimensionality = self.args.z_dimensions
		self.dataset = dataset

		# Global input size: trajectory at every step - x,y,action
		# Inputs is now states and actions.
		# Model size parameters
		self.state_size = 2
		self.state_dim = 2
		self.input_size = 2*self.state_size
		self.hidden_size = self.args.hidden_size
		# Number of actions
		self.output_size = 2					
		self.number_layers = self.args.number_layers
		self.traj_length = 5
		self.conditional_info_size = 6		

		if self.args.data=='MIME':
			self.state_size = 16	
			self.state_dim = 16
			self.input_size = 2*self.state_size
			self.output_size = self.state_size			
			self.traj_length = self.args.traj_length

			# Create Baxter visualizer for MIME data
			# self.visualizer = BaxterVisualizer.MujocoVisualizer()
			self.visualizer = BaxterVisualizer()

			if self.args.normalization=='meanvar':
				self.norm_sub_value = np.load("Statistics/MIME_Means.npy")
				self.norm_denom_value = np.load("Statistics/MIME_Var.npy")
			elif self.args.normalization=='minmax':
				self.norm_sub_value = np.load("Statistics/MIME_Min.npy")
				self.norm_denom_value = np.load("Statistics/MIME_Max.npy") - np.load("Statistics/MIME_Min.npy")

			# Max of robot_state + object_state sizes across all Baxter environments. 			
			self.cond_robot_state_size = 60
			self.cond_object_state_size = 25
			self.conditional_info_size = self.cond_robot_state_size+self.cond_object_state_size
			self.conditional_viz_env = False

		elif self.args.data=='Roboturk' or self.args.data=='OrigRoboturk' or self.args.data=='FullRoboturk':
			self.state_size = 8	
			self.state_dim = 8
			self.input_size = 2*self.state_size	
			self.output_size = self.state_size
			self.traj_length = self.args.traj_length

			self.visualizer = SawyerVisualizer()

			# Max of robot_state + object_state sizes across all sawyer environments. 
			# Robot size always 30. Max object state size is... 23. 
			self.cond_robot_state_size = 30
			self.cond_object_state_size = 23
			self.number_tasks = 8
			self.conditional_info_size = self.cond_robot_state_size+self.cond_object_state_size+self.number_tasks
			self.conditional_viz_env = True

		elif self.args.data=='Mocap':
			self.state_size = 22*3
			self.state_dim = 22*3	
			self.input_size = 2*self.state_size
			self.hidden_size = self.args.hidden_size
			self.output_size = self.state_size
			self.traj_length = self.args.traj_length	
			self.conditional_info_size = 0
			self.conditional_information = None
			self.conditional_viz_env = False

			# Create visualizer object
			self.visualizer = MocapVisualizer(args=self.args)

		self.training_phase_size = self.args.training_phase_size
		self.number_epochs = self.args.epochs
		self.test_set_size = 500
		self.baseline_value = 0.
		self.beta_decay = 0.9

		self. learning_rate = self.args.learning_rate

		self.latent_b_loss_weight = self.args.lat_b_wt
		self.latent_z_loss_weight = self.args.lat_z_wt

		self.initial_epsilon = self.args.epsilon_from
		self.final_epsilon = self.args.epsilon_to
		self.decay_epochs = self.args.epsilon_over
		self.decay_counter = self.decay_epochs*len(self.dataset)

		# Log-likelihood penalty.
		self.lambda_likelihood_penalty = self.args.likelihood_penalty
		self.baseline = None

		# Per step decay. 
		self.decay_rate = (self.initial_epsilon-self.final_epsilon)/(self.decay_counter)