def __init__(self,
n_population,
pc,
pm,
bankruptcy_data,
non_bankruptcy_data,
clusters_data,
cluster_centers,
threshold_list,
population=None):
self.threshold_list = threshold_list
self.bankruptcy_data = bankruptcy_data
self.non_bankruptcy_data = non_bankruptcy_data
self.neural_network = NeuralNetwork(n_inputs=6,
n_outputs=2,
n_neurons_to_hl=6,
n_hidden_layers=1)
self.n_population = n_population
self.p_crossover = pc # percent of crossover
self.p_mutation = pm # percent of mutation
self.population = population or self._makepopulation()
self.saved_cluster_data = clusters_data
self.cluster_centers = cluster_centers
self.predict_bankruptcy = []
self.predict_non_bankruptcy = []
self.fitness_list = [] # list of chromosome and fitness
self.currentUnderSampling = None
self.predict_chromosome = None
self.fitness()
def test_fit_and_predict(self):
ann = NeuralNetwork([4, 2], alpha=1e-5)
ann.fit(self.X, self.y)
T = self.X[[10, 60, 110]]
predictions = ann.predict(T)
print(predictions)
np.testing.assert_array_equal(predictions, np.array([0, 1, 2]))
def test_predict_probabilities(self):
ann = NeuralNetwork([4, 2], alpha=1e-5)
ann.fit(self.X, self.y)
T = self.X[[15, 65, 115, 117]]
ps = ann.predict_proba(T)
margin = np.min(np.max(ps, axis=1))
self.assertGreater(margin, 0.90)
def test_on_digits(self):
data_full = datasets.load_digits()
data, resp = utils.shuffle(data_full.data, data_full.target)
m = data.shape[0]
X, y = data[:m // 2], resp[:m // 2]
X_test, y_test = data[m // 2:], resp[m // 2:]
ann = NeuralNetwork([20, 5], alpha=1e-5)
ann.fit(X, y)
y_hat = ann.predict(X_test)
acc = metrics.accuracy_score(y_test, y_hat)
self.assertGreater(acc, 0.85)
def test_with_crossvalidation(self):
from sklearn.model_selection import cross_validate
clf = NeuralNetwork([10, 2], alpha=1e-5)
scores = cross_validate(clf, self.X, self.y, scoring='accuracy', cv=5)
acc = np.sum(scores["test_score"]) / 5
self.assertGreater(acc, 0.94)
def __init__(self, **kwargs):
self._quote_collection_name = kwargs.get("quote_collection")
self._qoute_mongo_client = ToolMongoClient(
kwargs.get("base_cfg_file", "mongo.conf"))
self._class_type = kwargs.get("class_type")
self._time_class = kwargs.get("time_class")
self.X, self.y = self.__get_data()
self.ann = NeuralNetwork([4, 3, 4], "tanh")
def test_gradient_computation(self):
ann = NeuralNetwork([2, 2], alpha=1e-5)
ann.set_data_(self.X, self.y)
coefs = ann.init_weights_()
g1 = ann.grad_approx(coefs, e=1e-5)
g2 = ann.grad(coefs)
np.testing.assert_array_almost_equal(g1, g2, decimal=10)
def evaluate_fitness_of_phenotype(cls, phenotype):
# Copy flatland so we can reuse for all phenotypes
flatland_scenarios = [
deepcopy(flatland_scenario)
for flatland_scenario in cls.flatland_scenarios
]
# Init neural network layers from phenotype weights
layers = list()
for layer_weight in phenotype.layer_weights:
layers.append(NeuronLayer(layer_weight))
# Init phenotype neural network
ann = NeuralNetwork(layers)
# Init phenotype agent
agent = FlatlandAgent(ann)
# Init fitness container used for avg computation
fitness_scenarios = list()
# Run agent for scenarios
for flatland_scenario in flatland_scenarios:
# Init variables for scenario fitness evaluation
phenotype_timesteps = 1
poisons = 0
foods = 0
while phenotype_timesteps != cls.max_time_steps:
# Get sensor data [left, front, right]
cells = flatland_scenario.get_sensible_cells()
# Let agent choose action based on sensor data
action = agent.choose_action(cells)
# Effect of action
if action != Move.STAND_STILL:
cell_value = cells[action.value - 1]
if cell_value == Flatland.food:
foods += 1
elif cell_value == Flatland.poison:
poisons += 1
# Commit action to world
flatland_scenario.move_agent(action)
phenotype_timesteps += 1
# Add fitness evaluation for scenario
fitness_scenarios.append(cls.fitness_function(foods, poisons))
# Evaluate fitness of agent and add it to collection
return sum(fitness_scenarios) / len(fitness_scenarios)
def test_weighs_structure(self):
ann = NeuralNetwork([5, 3], alpha=1e-5)
ann.set_data_(self.X, self.y)
coefs = ann.unflatten_coefs(ann.init_weights_())
shapes = np.array([coef.shape for coef in coefs])
np.testing.assert_array_equal(shapes, np.array([[5, 5], [6, 3], [4,
3]]))
def main():
size_of_learn_sample = int(len(x) * 0.9)
print(size_of_learn_sample)
NN = NeuralNetwork(x, y, 0.5)
# NN.print_matrices()
NN.train()
NN.print_matrices()
def test_ann(X_train, y_train, X_test, y_test, layers, max_iter=100):
""" """
nn = NeuralNetwork(layers, [ Activation() ] * (len(layers) - 1),
# regularization_rate=0.0001,
# regularization=Regularization('L2'),
cost=Cost('Bernoulli') )
# print layers, ', ', nn.neurons[1].type, \
# ', ', nn.regular.type, \
# ', ', nn.cost.type, '\n' #
cost, cv_cost = 10., 0.
accr, cv_accr = 0., 0.
cv_cost_prev = 0.
# tolerance = np.finfo(float).eps
cost_threshold = 0.1 # min_cost: np.finfo(float).eps
stop_threshold = 0.05
total_time = time.time()
start_time = time.time()
iter, n_iter_skip = 0, 100
# iter, max_iter, n_iter_skip = 0, 1000, 100
# while cost > cost_threshold:
while iter < max_iter:
Pt = nn.predict_matrix(X_test)
# cv_cost = nn.calculate_cost(X_test, y_test, Pt)
cv_accr = nn.accuracy(X_test, y_test, Pt)
P = nn.predict_matrix(X_train)
# cost = nn.calculate_cost(X_train, y_train, P)
accr = nn.accuracy(X_train, y_train, P)
if not (iter % n_iter_skip):
print '%4d\ttime=%.3f sec.\tCost(X=%.3f T=%.3f)\tAccr(X=%.3f T=%.3f)' % \
(iter, time.time() - start_time, cost, cv_cost, accr, cv_accr)
start_time = time.time()
# # Ранняя остановка на основании ошибки на валидационном множестве
# if cv_cost_prev and (cv_cost - cv_cost_prev) > stop_threshold:
# break
# else: cv_cost_prev = cv_cost
# if iter > max_iter: break
nn.backprop(X_train, y_train, P)
iter += 1
print '\n%4d\ttime=%.3f sec.\tCost(X=%.3f T=%.3f)\tAccr(X=%.3f T=%.3f)' % \
(iter, time.time() - total_time, cost, cv_cost, accr, cv_accr)
return accr, cv_accr # cost, cv_cost
def train(net: ann.NeuralNetwork,
samples: Samples,
error_function: Mapping[Tuple(Sample, ann.NeuralNetwork), float],
goal,
learning_rate=0.1,
epoches=10
) -> ann.NeuralNetwork:
""" Back propagation trainer.
Parameters
---------
goal:
epoches:
"""
result_net = net.copy()
for epoch in range(epoches):
for sample in samples:
reach_the_goal = abs(error_function(sample, result_net)) < goal
if reach_the_goal:
return result_net
else:
result_net = update(result_net, sample,
error_function, learning_rate
)
return result_net
from ann import NeuralNetwork
import matplotlib.pyplot as plt
import pickle
ground_truth_dataset = [[15, 3, 1], [10, 5, 1], [20, 1, 1],
[1, 5, 0], [5, 0, 0], [30, 0, 1], [2, 1, 0], [5, 5, 1],
[7, 10, 0], [25, 6, 1]]
n = NeuralNetwork()
# n.w1 = 0.024739923179843564
# n.w2 = 0.782086203350603
# n.w3 = -0.7531968177831182
# n.w4 = -0.1328632433367266
# n.w5 = 0.7521920897144088
# n.w6 = -0.2568310157160393
# n.b1 = -0.5874532717161877
# n.b2 = -0.9232275728505895
# n.b3 = 0.75530633226724
n.w1 = 0.27985946828331204
n.w2 = 1.113212641433306
n.w3 = -1.8082799142509993
n.w4 = 1.572218879949482
n.w5 = 4.689412543076026
n.w6 = -4.6976243908676985
n.b1 = -3.6516812511529557
n.b2 = -0.5814582597051287
n.b3 = -2.1295699805201465
print("""Initial hyperparameters:
n.w1 = {}
n.w2 = {}
def test_set_data(self):
ann = NeuralNetwork([5, 3], alpha=1e-5)
ann.set_data_(self.X, self.y)
coefs = ann.init_weights_()
self.assertEqual(len(coefs), 55)
import GA
import numpy as np
from ann import NeuralNetwork
sol_per_pop = 8
num_parents_mating = 4
crossover_location = 5
# Defining the population size.
pop_size = (sol_per_pop) # The population will have sol_per_pop chromosome where each chromosome has num_weights genes.
print(pop_size)
new_population = []
for i in range(sol_per_pop):
new_network = NeuralNetwork()
weights = []
#Input Layer
input_weights=np.random.rand(4, 6) #weight
input_biases=np.random.rand(6) #biases
weights.append(input_weights)
weights.append(input_biases)
#Hidden Layers
hidden_weights=np.random.rand(6, 6) #weight
hidden_biases=np.random.rand(6) #biases
weights.append(hidden_weights)
weights.append(hidden_biases)
#Output Layer
output_weights=np.random.rand(6, 3) #weight
output_biases=np.random.rand(3) #biases
weights.append(output_weights)
weights.append(output_biases)
parser.add_argument(
"--batchSize", help="Batch size count used for training data.", default=1024, type=int)
parser.add_argument("--dropout", help="Percent Dropout",
default=0.25, type=float)
parser.add_argument("--train", help="Train dataset", default='dataset/train', type=str)
parser.add_argument("--test", help="Test dataset", default='dataset/test', type=str)
if __name__ == "__main__":
args = parser.parse_args()
ann = NeuralNetwork(
tsSize=args.timeseries,
lstmSize=args.lstmSize,
dropout=args.dropout,
)
if os.path.isfile(args.weights):
ann.model.load_weights(args.weights)
if args.action == "train":
ann.fit(args.weights, args.train, args.test,
epochs=args.epochs, batch_size=args.batchSize)
if args.action == "serve":
serve(ann, args.test)
if args.action == "cli":
while 1:
def __init__(self):
self.network = NeuralNetwork()
self.fitness = -1
def generate_population(self):
self.population = []
for i in range(self.population_max_size):
agent = Agent(i, NeuralNetwork())
self.population.append(agent)
def __init__(self, flatland_scenarios, genotype_agent, time_steps, *args,
**kwargs):
Tk.__init__(self, *args, **kwargs)
self.configure(background="#2b2b2b")
# Static view state variables
self.max_time_steps = time_steps
self.phenotype_agent = genotype_agent.translate_to_phenotype()
self.flatland_agent = FlatlandAgent(
NeuralNetwork(
[NeuronLayer(w) for w in self.phenotype_agent.layer_weights]))
self.canvas = Canvas(self,
width=FlatlandView.viewport_width,
height=FlatlandView.viewport_height,
bg="#a0a0a0",
highlightbackground="#000000")
self.agent_polygon_x_y = [
10, -10, 30, 0, 10, 10, 0, 30, -10, 10, -30, 0, -10, -10
]
self.agent_senors_x_y = [
30, -10, 40, 0, 30, 10, 20, 0, 0, 40, -10, 30, 0, 20, 10, 30, -30,
-10, -20, 0, -30, 10, -40, 0
]
self.flatland_length = flatland_scenarios[0].length
self.agent_start = flatland_scenarios[0].agent_start
# Dynamic view state variables
self.flatland_scenarios = flatland_scenarios
self.current_scenario = 0
self.current_flatland_scenario = deepcopy(
self.flatland_scenarios[self.current_scenario])
self.time_steps = 0
self.time_step_delay = IntVar(self, 1000)
self.time_step_stringvar = StringVar(self, str(self.time_steps))
self.sensor_rotation = 0
# Draw flatland grid and configure canvas
self.draw_canvas_lines()
self.canvas.grid(rowspan=18, columnspan=3)
self.time_step_num_font = tkFont.Font(family="Helvetica",
size=72,
weight="bold")
self.time_step_text_font = tkFont.Font(family="Helvetica",
size=18,
weight="bold")
self.time_step_frame = Frame(self, background="#2b2b2b")
Label(self.time_step_frame,
textvariable=self.time_step_stringvar,
font=self.time_step_num_font,
background="#2b2b2b",
foreground="#a9b7c6").pack()
Label(self.time_step_frame,
text="time steps",
font=self.time_step_text_font,
background="#2b2b2b",
foreground="#a9b7c6").pack()
self.time_step_frame.grid(row=0, column=4)
Scale(self,
from_=250,
to=5000,
resolution=250,
background="#2b2b2b",
foreground="#a9b7c6",
label="Time step delay (ms)",
variable=self.time_step_delay).grid(row=16, column=4, padx=20)
Button(self,
text="New Scenarios",
width=25,
command=self.generate_new_scenarios,
foreground="#a9b7c6",
background="#2b2b2b",
highlightbackground="#2b2b2b").grid(row=17, column=4, padx=20)
class GA(object):
def __init__(self,
n_population,
pc,
pm,
bankruptcy_data,
non_bankruptcy_data,
clusters_data,
cluster_centers,
threshold_list,
population=None):
self.threshold_list = threshold_list
self.bankruptcy_data = bankruptcy_data
self.non_bankruptcy_data = non_bankruptcy_data
self.neural_network = NeuralNetwork(n_inputs=6,
n_outputs=2,
n_neurons_to_hl=6,
n_hidden_layers=1)
self.n_population = n_population
self.p_crossover = pc # percent of crossover
self.p_mutation = pm # percent of mutation
self.population = population or self._makepopulation()
self.saved_cluster_data = clusters_data
self.cluster_centers = cluster_centers
self.predict_bankruptcy = []
self.predict_non_bankruptcy = []
self.fitness_list = [] # list of chromosome and fitness
self.currentUnderSampling = None
self.predict_chromosome = None
self.fitness()
def init_neural_network(self, chromosome):
# remove threshold from chromosome list
primary_weights = chromosome[5:]
matrix_list = []
for i in range(0, int(len(primary_weights) / 6) - 1):
matrix_list.append(primary_weights[i * 6:(i + 1) * 6])
weights_matrix = array(matrix_list)
layers = self.neural_network.layers
i = 0
for neuron in layers[0].neurons:
neuron.set_weights(weights_matrix[:, i])
i += 1
layers[1].neurons[0].set_weights(primary_weights[-6:])
def performance_measure(self):
tp = 0
fp = 0
fn = 0
tn = 0
for item in self.bankruptcy_data:
if self.predict(item) > 0.5:
fp += 1
else:
tp += 1
for item in self.non_bankruptcy_data:
if self.predict(item) > 0.5:
tn += 1
else:
fn += 1
sensitivity = tp / (tp + fn)
specificity = tn / (fp + tn)
print("TP is : %s" % (str(tp)))
print("FP is : %s" % (str(fp)))
print("FN is : %s" % (str(fn)))
print("TN is : %s" % (str(tn)))
print("G-MEAN : %s" % (str(math.sqrt(sensitivity * specificity))))
print("Hit-ratio : %s" % (str((tp + tn) / (tp + fn + fp + tn))))
def predict(self, data):
self.init_neural_network(self.predict_chromosome)
return self.neural_network.update(data)[0]
def _makepopulation(self):
pop_list = []
for i in range(0, self.n_population):
weights = [random.uniform(-5, 5) for _ in range(0, 36)]
out_weights = [random.uniform(-5, 5) for _ in range(0, 12)]
# make threshold list
threshold1 = [
random.uniform(threshold[0], threshold[1])
for threshold in self.threshold_list
]
chromosome = threshold1 + weights + out_weights
pop_list.append(chromosome)
return pop_list
'''
b : the number of bankruptcy firms
BAi : the classification accuracy of ith instances of bankruptcy firms
n : the number of non-bankruptcy firms
NAj : the classification accuracy of jth instances of non-bankruptcy firms
POi : the predicated output of ith instances of bankruptcy firms
AOi : the actual output of ith instances of non-bankruptcy firms
POj : the predicated output of jth instances of non-bankruptcy firms
AOj : the actual output of jth instances of non-bankruptcy firms
'''
def ba_i(self, poi):
if poi < 0.5:
return 1
return 0
def na_j(self, poj):
if poj > 0.5:
return 1
return 0
def cbeus(
self, thresholds
): # the rule structure for the cluster-based underSampling base on GA
i = 0
undersampling_clusters = []
for cluster in self.saved_cluster_data:
for instance in cluster:
if euclidean_distances(
[instance], [self.cluster_centers[i]]) < thresholds[i]:
undersampling_clusters.append(instance)
i += 1
return undersampling_clusters
def fitness(self):
fitness_sum = 0
trials = 3
for index in range(0, trials):
for item in self.population:
print("underSampling : Cut off % s" % str(item[:5]))
self.currentUnderSampling = self.cbeus(item[:5])
self.init_neural_network(item)
self.predict_non_bankruptcy = []
self.predict_bankruptcy = []
for instance in self.currentUnderSampling:
self.predict_non_bankruptcy.append(
self.neural_network.update(instance)[0])
for instance in self.bankruptcy_data:
self.predict_bankruptcy.append(
self.neural_network.update(instance)[0])
fitness_value = self.g_mean(len(self.currentUnderSampling))
self.fitness_list.append([item, fitness_value])
fitness_sum += fitness_value
self.fitness_list.sort(key=lambda x: x[1])
self._select_parents(fitness_sum)
self.fitness_list.sort(key=lambda x: x[1])
self.population = []
for item in self.fitness_list:
self.population.append(item[0])
if len(self.population) == 5:
break
if index == trials - 1:
os.system('cls' if os.name == 'nt' else 'clear')
print("The Optimization Weights For Predict Is: %s " %
str(self.population[0][5:]))
self.predict_chromosome = self.population[0][5:]
self.performance_measure()
def g_mean(self, n):
b = len(self.bankruptcy_data)
sum_bankruptcy = 0
sum_non_bankruptcy = 0
for item in self.predict_non_bankruptcy:
sum_non_bankruptcy += self.ba_i(item)
for item in self.predict_bankruptcy:
sum_bankruptcy += self.na_j(item)
return math.sqrt(
(1 / b) * sum_bankruptcy * (1 / n) * sum_non_bankruptcy)
def cxOnePoint(self, ind1, ind2):
"""Executes a one point crossover on the input :term:`sequence` individuals.
The two individuals are modified in place. The resulting individuals will
respectively have the length of the other.
:param ind1: The first individual participating in the crossover.
:param ind2: The second individual participating in the crossover.
:returns: A tuple of two individuals.
This function uses the :func:`~random.randint` function from the
python base :mod:`random` module.
"""
size = min(len(ind1), len(ind2))
cxpoint = random.randint(1, size - 1)
ind1[cxpoint:], ind2[cxpoint:] = ind2[cxpoint:], ind1[cxpoint:]
return ind1, ind2
def swapMutation(self, ind1):
size = len(ind1)
swpoint1 = random.randint(1, size - 1)
swpoint2 = random.randint(1, size - 1)
ind1[swpoint1], ind1[swpoint2] = ind1[swpoint2], ind1[swpoint1]
return ind1
def _select_parents(self, fitness_sum):
"""
Roulette wheel selection
Selects parents from the given population
Args :
population (list) : Current population from which parents will be selected
fitness_sum (number) : Summation of all fitness value
Returns :
parents (IndividualGA, IndividualGA) : selected parents
"""
probability = []
for item in self.fitness_list:
probability.append(item[1] / fitness_sum)
item.append(item[1] / fitness_sum)
ncrossover = math.ceil(self.n_population * self.p_crossover /
2) # number of crossover offspring
nmutation = math.ceil(self.n_population *
self.p_mutation) # number of mutation offspring
selection_probability = set()
while len(selection_probability) < ncrossover:
selection_probability.add(random.uniform(0, 1))
probability = np.cumsum(probability).tolist()
def roulette(prob):
for i in range(0, len(probability)):
if prob < probability[i]:
return self.fitness_list[i][0]
crossover_list = []
for item in list(selection_probability):
crossover_list.append(roulette(item))
if len(crossover_list) == 2:
inde1, inde2 = self.cxOnePoint(crossover_list[0][:],
crossover_list[1][:])
# init the neural network with the individual 1
self.currentUnderSampling = self.cbeus(inde1[:5])
self.init_neural_network(inde1)
self.predict_non_bankruptcy = []
self.predict_bankruptcy = []
for instance in self.currentUnderSampling:
self.predict_non_bankruptcy.append(
self.neural_network.update(instance)[0])
for instance in self.bankruptcy_data:
self.predict_bankruptcy.append(
self.neural_network.update(instance)[1])
fitness_value = self.g_mean(len(self.currentUnderSampling))
self.fitness_list.append([inde1, fitness_value])
# init the neural network with the individual 2
self.currentUnderSampling = self.cbeus(inde2[:5])
self.init_neural_network(inde2)
self.predict_non_bankruptcy = []
self.predict_bankruptcy = []
for instance in self.currentUnderSampling:
self.predict_non_bankruptcy.append(
self.neural_network.update(instance)[0])
for instance in self.bankruptcy_data:
self.predict_bankruptcy.append(
self.neural_network.update(instance)[1])
fitness_value = self.g_mean(len(self.currentUnderSampling))
self.fitness_list.append([inde2, fitness_value])
crossover_list = []
# create individual with mutation
selection_probability = set()
while len(selection_probability) < nmutation:
selection_probability.add(random.uniform(0, 1))
for item in list(selection_probability):
inde3 = self.swapMutation(roulette(item))
self.currentUnderSampling = self.cbeus(inde3[:5])
self.init_neural_network(inde3)
self.predict_non_bankruptcy = []
self.predict_bankruptcy = []
for instance in self.currentUnderSampling:
self.predict_non_bankruptcy.append(
self.neural_network.update(instance)[0])
for instance in self.bankruptcy_data:
self.predict_bankruptcy.append(
self.neural_network.update(instance)[1])
fitness_value = self.g_mean(len(self.currentUnderSampling))
self.fitness_list.append([inde3, fitness_value])
import pygame
import pickle
import numpy as np
from game import init, iterate
from ann import NeuralNetwork
import utils
# Architecture (Specify archetecture here.)
network = NeuralNetwork(layers=[7, 14, 14, 7, 1],
activations=['sigmoid', 'sigmoid', 'sigmoid', 'tanh'])
lr = 0.1
losses = []
screen, font = init()
# Game Loop / Train Loop
frame_count, score, _, _, x = iterate.iterate(screen, font, 0, 0)
game = True
run = True
prediction = 0
while run:
for event in pygame.event.get():
if event.type == pygame.QUIT:
run = False
prediction = utils.forward(x, network)
frame_count, score, game, run, x = iterate.iterate(screen, font,
frame_count, score,
game, run, prediction)
loss = utils.backward(prediction, x, lr, network)
losses.append(loss)
pygame.quit()
import pickle
ground_truth_dataset = [
[15, 3, 1],
[10, 5, 1],
[20, 1, 1],
[1, 5, 0],
[5, 0, 0],
[30, 0, 1],
[2, 1, 0],
[5, 5, 1],
[7, 10, 0],
[25, 6, 1]
]
n = NeuralNetwork()
stats = n.train(ground_truth_dataset, 100001, 0.001)
epochs = stats[0]
min_losses = stats[1]
avg_losses = stats[2]
max_losses = stats[3]
with open("model.bin", "wb") as f:
pickle.dump(n, f)
plt.ylabel("Loss")
plt.xlabel("Epoch")
plt.plot(epochs, min_losses, label="Min loss")
plt.plot(epochs, avg_losses, label="Avg loss")
plt.plot(epochs, max_losses, label="Max loss")
class GWO(object):
def __init__(self,
n_population,
bankruptcy_data,
non_bankruptcy_data,
clusters_data,
cluster_centers,
threshold_list,
population=None):
self.threshold_list = threshold_list
self.bankruptcy_data = bankruptcy_data
self.non_bankruptcy_data = non_bankruptcy_data
self.neural_network = NeuralNetwork(n_inputs=6,
n_outputs=2,
n_neurons_to_hl=6,
n_hidden_layers=1)
self.n_population = n_population
self.population = population or self._makepopulation()
self.saved_cluster_data = clusters_data
self.cluster_centers = cluster_centers
self.predict_bankruptcy = []
self.predict_non_bankruptcy = []
self.fitness_list = [] # list of chromosome and fitness
self.currentUnderSampling = None
self.predict_position = None
self.search_main()
def init_neural_network(self, chromosome):
# remove threshold from chromosome list
primary_weights = chromosome[5:]
matrix_list = []
for i in range(0, int(len(primary_weights) / 6) - 1):
matrix_list.append(primary_weights[i * 6:(i + 1) * 6])
weights_matrix = array(matrix_list)
layers = self.neural_network.layers
i = 0
for neuron in layers[0].neurons:
neuron.set_weights(weights_matrix[:, i])
i += 1
layers[1].neurons[0].set_weights(primary_weights[-6:])
def performance_measure(self):
tp = 0
fp = 0
fn = 0
tn = 0
for item in self.bankruptcy_data:
if self.predict(item) > 0.5:
fp += 1
else:
tp += 1
for item in self.non_bankruptcy_data:
if self.predict(item) > 0.5:
tn += 1
else:
fn += 1
sensitivity = tp / (tp + fn)
specificity = tn / (fp + tn)
print("TP is : %s" % (str(tp)))
print("FP is : %s" % (str(fp)))
print("FN is : %s" % (str(fn)))
print("TN is : %s" % (str(tn)))
print("G-MEAN : %s" % (str(math.sqrt(sensitivity * specificity))))
print("Hit-ratio : %s" % (str((tp + tn) / (tp + fn + fp + tn))))
def predict(self, data):
self.init_neural_network(self.predict_position)
return self.neural_network.update(data)[0]
def _makepopulation(self):
pop_list = []
for i in range(0, self.n_population):
weights = [random.uniform(-5, 5) for _ in range(0, 36)]
out_weights = [random.uniform(-5, 5) for _ in range(0, 12)]
# make threshold list
threshold1 = [
random.uniform(threshold[0], threshold[1])
for threshold in self.threshold_list
]
position = threshold1 + weights + out_weights
pop_list.append(position)
return pop_list
def ba_i(self, poi):
if poi < 0.5:
return 1
return 0
def na_j(self, poj):
if poj > 0.5:
return 1
return 0
def cbeus(
self, thresholds
): # the rule structure for the cluster-based underSampling base on GA
i = 0
undersampling_clusters = []
for cluster in self.saved_cluster_data:
for instance in cluster:
if euclidean_distances(
[instance], [self.cluster_centers[i]]) < thresholds[i]:
undersampling_clusters.append(instance)
i += 1
return undersampling_clusters
def search_main(self):
Max_iter = 3
for index in range(0, Max_iter):
for position in self.population:
print("underSampling : Cut off % s" % str(position[:5]))
self.currentUnderSampling = self.cbeus(position[:5])
self.init_neural_network(position)
self.predict_non_bankruptcy = []
self.predict_bankruptcy = []
for instance in self.currentUnderSampling:
self.predict_non_bankruptcy.append(
self.neural_network.update(instance)[0])
for instance in self.bankruptcy_data:
self.predict_bankruptcy.append(
self.neural_network.update(instance)[0])
fitness_value = self.g_mean(len(self.currentUnderSampling))
self.fitness_list.append([position, fitness_value])
self.fitness_list.sort(key=lambda x: x[1])
# Update Alpha, Beta, and Delta
Alpha_pos = self.fitness_list[0][0] # Update alpha
Beta_pos = self.fitness_list[1][0] # Update beta
Delta_pos = self.fitness_list[2][0] # Update delta
a = 2 - index * (
(2) / Max_iter) # a decreases linearly from 2 to 0
for position in self.population:
for j in range(0, len(position)):
r1 = random.random() # r1 is a random number in [0,1]
r2 = random.random() # r2 is a random number in [0,1]
A1 = 2 * a * r1 - a # Equation (3.3)
C1 = 2 * r2 # Equation (3.4)
D_alpha = abs(C1 * Alpha_pos[j] -
position[j]) # Equation (3.5)-part 1
X1 = Alpha_pos[j] - A1 * D_alpha # Equation (3.6)-part 1
r1 = random.random()
r2 = random.random()
A2 = 2 * a * r1 - a # Equation (3.3)
C2 = 2 * r2 # Equation (3.4)
D_beta = abs(C2 * Beta_pos[j] -
position[j]) # Equation (3.5)-part 2
X2 = Beta_pos[j] - A2 * D_beta # Equation (3.6)-part 2
r1 = random.random()
r2 = random.random()
A3 = 2 * a * r1 - a # Equation (3.3)
C3 = 2 * r2 # Equation (3.4)
D_delta = abs(C3 * Delta_pos[j] -
position[j]) # Equation (3.5)-part 3
X3 = Delta_pos[j] - A3 * D_delta # Equation (3.5)-part 3
position[j] = (X1 + X2 + X3) / 3 # Equation (3.7)
if index == Max_iter - 1:
os.system('cls' if os.name == 'nt' else 'clear')
self.fitness_list.sort(key=lambda x: x[1])
print("The Optimization Weights For Predict Is: %s " %
str(self.fitness_list[0][0][5:]))
self.predict_position = self.fitness_list[0][0][5:]
self.performance_measure()
def g_mean(self, n):
b = len(self.bankruptcy_data)
sum_bankruptcy = 0
sum_non_bankruptcy = 0
for item in self.predict_non_bankruptcy:
sum_non_bankruptcy += self.ba_i(item)
for item in self.predict_bankruptcy:
sum_bankruptcy += self.na_j(item)
return math.sqrt(
(1 / b) * sum_bankruptcy * (1 / n) * sum_non_bankruptcy)