def __init__()

in Experiments/PolicyNetworks.py [0:0]

• 30 lines of code
• 9 McCabe index (conditional complexity)

	def __init__(self, input_size, hidden_size, args, number_layers=4):		

		# Ensures inheriting from torch.nn.Module goes nicely and cleanly. 	
		# super().__init__()
		super(ContinuousLatentPolicyNetwork, self).__init__()

		self.args = args
		# Input size is actually input_size + number_subpolicies +1 
		self.input_size = input_size+self.args.z_dimensions+1
		self.offset_for_z = input_size+1
		self.hidden_size = hidden_size
		# self.number_subpolicies = number_subpolicies
		self.output_size = self.args.z_dimensions
		self.num_layers = number_layers
		self.b_exploration_bias = self.args.b_exploration_bias
		self.batch_size = self.args.batch_size

		# Define LSTM. 
		self.lstm = torch.nn.LSTM(input_size=self.input_size,hidden_size=self.hidden_size,num_layers=self.num_layers).to(device)

		# Transform to output space - Latent z and Latent b. 
		# self.subpolicy_output_layer = torch.nn.Linear(self.hidden_size,self.output_size)
		self.termination_output_layer = torch.nn.Linear(self.hidden_size,2)
		
		# Sigmoid and Softmax activation functions for Bernoulli termination probability and latent z selection .
		self.batch_softmax_layer = torch.nn.Softmax(dim=2)
		self.batch_logsoftmax_layer = torch.nn.LogSoftmax(dim=2)

		# Define output layers for the LSTM, and activations for this output layer. 
		self.mean_output_layer = torch.nn.Linear(self.hidden_size,self.output_size)
		self.variances_output_layer = torch.nn.Linear(self.hidden_size, self.output_size)

		self.activation_layer = torch.nn.Tanh()
		self.variance_activation_layer = torch.nn.Softplus()
		self.variance_activation_bias = 0.
			
		self.variance_factor = 0.01

		# # # Try initializing the network to something, so that we can escape the stupid constant output business.
		for name, param in self.lstm.named_parameters():
			if 'bias' in name:
				torch.nn.init.constant_(param, 0.001)
			elif 'weight' in name:
				torch.nn.init.xavier_normal_(param,gain=5)

		# Also initializing mean_output_layer to something large...
		for name, param in self.mean_output_layer.named_parameters():
			if 'bias' in name:
				torch.nn.init.constant_(param, 0.)
			elif 'weight' in name:
				torch.nn.init.xavier_normal_(param,gain=2)