def test(self, inputs, outputs, fitness_callback, to_test='gen_best'):
if self.network_params['network'] == 'feedforward':
model = Network(0, self.gen_best)
with tf.Session() as sess:
actions = sess.run(model.prediction,
feed_dict={model.X: inputs})
print('test score: {0:.2f}%'.format(
fitness_callback(actions, outputs)))
tf.reset_default_graph()
# built using tf.keras
else:
# reshape data for each network type
if self.network_params['network'] == 'convolutional':
inputs = inputs[:, :, np.newaxis]
elif self.network_params['network'] == 'recurrent':
timesteps = self.network_params['timesteps']
inputs = np.array([
inputs[i - timesteps:i]
for i in range(timesteps, inputs.shape[0] + 1)
])
outputs = outputs[timesteps - 1:]
actions = self.gen_best.model.prediction.predict(inputs)
print('test score: {0:.2f}%'.format(
fitness_callback(actions, outputs)))
def test(self, inputs, outputs, fitness_callback, to_test='gen_best'):
model = Network(0, self.gen_best)
with tf.Session() as sess:
actions = sess.run(model.prediction, feed_dict={model.X: inputs})
print('test score: {0:.2f}%\t(hold: {1:.2f}%)'.format(
fitness_callback(actions, outputs),
(outputs[-1] / outputs[0] - 1) * 100))
tf.reset_default_graph()
def __init__(self,
id,
network_params,
mutation_scale,
w_mutation_rate,
b_mutation_rate,
parent_1=None,
parent_2=None):
# fitness is score normalized
self.fitness = 0
self.actions = None
self.prices = []
self.score = 0
self.id = id
# keep track of how genome came to be
self.mutated = False
self.bred = False
self.inputs = network_params['input']
self.hidden = network_params['hidden']
self.outputs = network_params['output']
self.network = network_params['network']
self.w_mutation_rate = w_mutation_rate
self.b_mutation_rate = b_mutation_rate
self.mutation_scale = mutation_scale
self.weights = None
self.biases = None
# two parents available for breeding
if parent_2 is not None:
self.weights = copy.deepcopy(parent_1.weights)
self.biases = copy.deepcopy(parent_1.biases)
self.breed(parent_2)
self.bred = True
# mutate if only one parent available
elif parent_1 is not None:
self.weights = copy.deepcopy(parent_1.weights)
self.biases = copy.deepcopy(parent_1.biases)
self.mutate_w_feedforward()
self.mutate_b_feedforward()
self.mutated = True
# initial values when population is first created
else: self.init_w_b()
self.model = Network(self.id, self)
def test(self, inputs, outputs, fitness_callback, to_test='gen_best'):
# built using tf.keras
if self.network_params['network'] == 'convolutional':
# add extra dimension for conv1D channel
inputs = inputs[:,:, np.newaxis]
actions = self.gen_best.model.prediction.predict(inputs)
print('test score: {0:.2f}%'.format(fitness_callback(actions, outputs)))
else:
model = Network(self.gen_best)
with tf.Session() as sess:
actions = sess.run(model.prediction, feed_dict={model.X: inputs})
print('test score: {0:.2f}%'.format(fitness_callback(actions, outputs)))
tf.reset_default_graph()
def __init__(self,
id,
network_params,
mutation_scale,
w_mutation_rate,
b_mutation_rate,
parent_1=None,
parent_2=None,
load_keras=None):
# fitness is score normalized
self.fitness = 0
self.score = 0
self.id = id
# keep track of how genome came to be
self.mutated = False
self.bred = False
self.inputs = network_params['input']
self.hidden = network_params['hidden']
self.outputs = network_params['output']
self.network = network_params['network']
self.model = None
self.timesteps = None
try:
if self.network == 'recurrent':
self.timesteps = network_params['timesteps']
except:
raise AttributeError(
'Must specify timesteps for recurrent network')
self.w_mutation_rate = w_mutation_rate
self.b_mutation_rate = b_mutation_rate
self.mutation_scale = mutation_scale
self.weights = None
self.biases = None
# two parents available for breeding
if parent_2 is not None:
self.weights = copy.deepcopy(parent_1.weights)
self.biases = copy.deepcopy(parent_1.biases)
self.breed(parent_2)
self.bred = True
# mutate if only one parent available
elif parent_1 is not None:
self.weights = copy.deepcopy(parent_1.weights)
self.biases = copy.deepcopy(parent_1.biases)
self.mutate_w_feedforward()
self.mutate_b_feedforward()
self.mutated = True
elif load_keras is not None:
self.model = Network(self.id, self, load_keras=load_keras)
self.mutate_w_cnn()
# initial values when population is first created
else:
self.init_w_b()
# pass genome to network object
if load_keras is None: self.model = Network(self.id, self)