def register_observations(self,space,new_center=None):
creator=self.observation_creator
if type(space) is tuple:
num_conv_layers = len([0 for layer in space if
layer[0] == 0])
self.current_num_layer=num_conv_layers
if new_center:
self.center_key=new_center
else:
self.center_key=space
else:
if new_center:
self.center_key=new_center
else:
self.center_key=space.center
center=space.center
self.current_num_layer=len([0 for layer in new_center if
layer[0] == 0])
node_c=space.key2node(center)
self.observations = []
weight_array = []
i = 0
for node in node_c.kids:
if Funct.node2energy(node) > 0:
weight_array.append(Funct.node2energy(node)*10)
weight_array = Funct.normalize(dataset=weight_array)
for node in node_c.kids:
if Funct.node2energy(node) > 0:
observation_path = []
for direction in self.current_path.path:
observation_path.append(direction)
absolute_direction = Funct.node2direction(node)
relative_direction = Funct.absolute2relative(direction=absolute_direction, last_layer=self.last_mutated_layer,
max_layers=self.max_layers_condition)
observation_path.append(relative_direction)
if len(observation_path) > self.max_observation_size:
observation_path.pop(0)
observation = self.observation_creator(path=observation_path, weight=weight_array[i])
self.observations.append(observation)
i += 1
def test():
observation_size = 3
max_layers_lstm = 5 * 2
num_actions = 4
mutations = ((0, 1, 0, 0), (0, -1, 0, 0), (0, 0, 1, 1), (0, 0, -1, -1))
directions_per_observations = [[(1, (0, 1, 0, 0)), (2, (0, -1, 0, 0)),
(0, (0, 1, 0, 0))],
[(1, (0, 1, 0, 0)), (2, (0, -1, 0, 0)),
(0, (0, 0, 1, 1))],
[(1, (0, 1, 0, 0)), (2, (0, -1, 0, 0)),
(0, (0, -1, 0, 0))],
[(1, (0, 1, 0, 0)), (2, (0, -1, 0, 0)),
(0, (0, 0, -1, -1))]]
#observations_weight = [getLoss(0.4525), getLoss(0.4520), getLoss(0.4145), getLoss(0.4220)]
observations_weight = [0.5142 * 10, 0.4821 * 10, 0.4680 * 10, 0.4734 * 10]
observations_weight = Funct.normalize(dataset=observations_weight)
print("weight: ", observations_weight)
lstmConverter = LSTMConverter.LSTMConverter(
cuda=True,
max_layers=max_layers_lstm,
mutation_list=mutations,
limit_directions=observation_size)
observations = []
for i in range(num_actions):
observation = Observation.Observation(
path=directions_per_observations[i],
weight=observations_weight[i],
time=0)
print("observation: ", i)
print("path: ", observation.path)
print("weight: ", observation.weight)
observations.append(observation)
input_lstm = lstmConverter.generate_LSTM_input(observations=observations)
print("input size: ", input_lstm.size())
current_path = Observation.Observation(path=[(1, (0, 1, 0, 0)),
(2, (0, -1, 0, 0))],
weight=1,
time=0)
network = nw_lstm.NetworkLSTM(observation_size=observation_size,
in_channels=max_layers_lstm,
out_channels=max_layers_lstm *
lstmConverter.mutations,
kernel_size=lstmConverter.mutations,
cuda_flag=True)
print("empezando entrenamiento")
network.Training(data=input_lstm,
dt=0.001,
p=5000,
observations=observations)
print("entrenamiento finalizado")
predict = lstmConverter.generate_LSTM_predict(observation=current_path)
predicted = network.predict(predict)
print("ht: ")
print(predicted)
predicted_directions = lstmConverter.topK_predicted_directions(
predicted_tensor=predicted, k=num_actions // 2)
print(predicted_directions)
i += 1