def cost_accuracy_string(net: Network, training: Batch, validation: Batch):
Y = net.evaluate(training.images)
accuracy = net.accuracy(training.one_hot_labels, outputs=Y)
cost = net.cost(training.one_hot_labels, outputs=Y)
res = 'Training:\n' + \
'- accuracy: {:.3%}\n'.format(accuracy) + \
'- cost: {:.3f}\n'.format(cost)
Y = net.evaluate(validation.images)
accuracy = net.accuracy(validation.one_hot_labels, outputs=Y)
cost = net.cost(validation.one_hot_labels, outputs=Y)
res += 'Validation:\n' + \
'- accuracy: {:.3%}\n'.format(accuracy) + \
'- cost: {:.3f}\n'.format(cost)
return res
def load_population(self, data):
pop = []
for n in data:
new_network = Network(n['nn_param_choices'])
new_network.network['nb_neurons'] = []
new_network.network['activation'] = []
new_network.network['nb_neurons'].extend(n['network']['nb_neurons'])
new_network.network['nb_layers'] = n['network']['nb_layers']
new_network.network['activation'].extend(n['network']['activation'])
new_network.network['optimizer'] = n['network']['optimizer']
new_network.accuracy = n['accuracy']
pop.append(new_network)
for c in pop:
if not self.check_network(c.network):
raise ValueError('Network out of balance')
return pop
class Model(object):
def __init__(self, config):
self.epoch_count = 0
self.config = config
self.data = DataSet(config)
self.add_placeholders()
self.summarizer = tf.summary
self.net = Network(config, self.summarizer)
self.optimizer = self.config.solver.optimizer
self.y_pred = self.net.prediction(self.x, self.keep_prob)
self.loss = self.net.loss(self.y_pred, self.y)
self.accuracy = self.net.accuracy(tf.nn.sigmoid(self.y_pred), self.y)
self.summarizer.scalar("accuracy", self.accuracy)
self.summarizer.scalar("loss", self.loss)
self.train = self.net.train_step(self.loss)
self.saver = tf.train.Saver()
self.init = tf.global_variables_initializer()
self.local_init = tf.local_variables_initializer()
def add_placeholders(self):
self.x = tf.placeholder(tf.float32, shape=[None, self.config.features_dim])
self.y = tf.placeholder(tf.float32, shape=[None, self.config.labels_dim])
self.keep_prob = tf.placeholder(tf.float32)
def run_epoch(self, sess, data, summarizer, epoch):
err = list()
i = 0
X, Y, Y_pred = None, None, None
merged_summary = self.summarizer.merge_all()
for X, Y, tot in self.data.next_batch(data):
feed_dict = {self.x : X, self.y : Y, self.keep_prob : self.config.solver.dropout}
if not self.config.load:
summ, _, loss, Y_pred = sess.run([merged_summary, self.train, self.loss, tf.nn.sigmoid(self.y_pred)], feed_dict=feed_dict)
err.append(loss)
output = "Epoch ({}) Batch({}) : Loss = {}".format(self.epoch_count, i , loss)
with open("../stdout/{}_train.log".format(self.config.project_name), "a+") as log:
log.write(output + "\n")
print(" {}".format(output), end='\r')
step = int(epoch*tot + i)
summarizer.add_summary(summ, step)
i += 1
err = np.asarray(err)
return np.mean(err), step
def run_eval(self, sess, data, summary_writer=None, step=0):
y, y_pred, loss_ = list(), list(), 0.0
loss = 0.0
merged_summary = self.summarizer.merge_all()
i = 0
for X, Y, tot in self.data.next_batch(data):
feed_dict = {self.x: X, self.y: Y, self.keep_prob: 1.0}
if i == tot-1 and summary_writer is not None:
summ, loss_ = sess.run([merged_summary, self.loss], feed_dict=feed_dict)
summary_writer.add_summary(summ, step)
else:
loss_ = sess.run(self.loss, feed_dict=feed_dict)
loss += loss_
i += 1
return loss / i
def get_metrics(self, sess, data):
accuracy, y_pred, i = 0.0, None, 0.0
for X, Y, tot in self.data.next_batch(data):
feed_dict = {self.x: X, self.y: Y, self.keep_prob: 1.0}
Y_pred, accuracy_val = sess.run([tf.nn.sigmoid(self.y_pred), self.accuracy], feed_dict=feed_dict)
metrics = evaluate(predictions=Y_pred, labels=Y)
p_k = patk(predictions=Y_pred, labels=Y)
accuracy += accuracy_val
i += 1
return metrics, accuracy / i, p_k
def add_summaries(self, sess):
if self.config.load or self.config.debug:
path_ = "../results/tensorboard"
else :
path_ = "../bin/results/tensorboard"
summary_writer_train = tf.summary.FileWriter(path_ + "/train", sess.graph)
summary_writer_val = tf.summary.FileWriter(path_ + "/val", sess.graph)
summary_writer_test = tf.summary.FileWriter(path_+ "/test", sess.graph)
summary_writers = {'train': summary_writer_train, 'val': summary_writer_val, 'test': summary_writer_test}
return summary_writers
def fit(self, sess, summarizer):
'''
- Patience Method :
+ Train for particular no. of epochs, and based on the frequency, evaluate the model using validation data.
+ If Validation Loss increases, decrease the patience counter.
+ If patience becomes less than a certain threshold, devide learning rate by 10 and switch back to old model
+ If learning rate is lesser than a certain
'''
sess.run(self.init)
sess.run(self.local_init)
max_epochs = self.config.max_epochs
patience = self.config.patience
patience_increase = self.config.patience_increase
improvement_threshold = self.config.improvement_threshold
best_validation_loss = 1e6
self.epoch_count = 0
best_step, losses, learning_rate = -1, list(), self.config.solver.learning_rate
while self.epoch_count < max_epochs :
if(self.config.load == True):
break
start_time = time.time()
average_loss, tr_step = self.run_epoch(sess, "train", summarizer['train'], self.epoch_count)
duration = time.time() - start_time
if not self.config.debug:
if self.epoch_count % self.config.epoch_freq == 0 :
val_loss = self.run_eval(sess, "validation", summarizer['val'], tr_step)
test_loss = self.run_eval(sess, "test", summarizer['test'], tr_step)
metrics, _, _ = self.get_metrics(sess, "test")
output = "=> Training : Loss = {:.3f} | Validation : Loss = {:.3f} | Test : Loss = {:.3f}".format(average_loss, val_loss, test_loss)
with open("../stdout/validation.log", "a+") as f:
output_ = output + "\n=> Test : Coverage = {}, Average Precision = {}, Micro Precision = {}, Micro Recall = {}, Micro F Score = {}".format(metrics['coverage'], metrics['average_precision'], metrics['micro_precision'], metrics['micro_recall'], metrics['micro_f1'])
output_ += "\n=> Test : Macro Precision = {}, Macro Recall = {}, Macro F Score = {}\n\n\n".format(metrics['macro_precision'], metrics['macro_recall'], metrics['macro_f1'])
f.write(output_)
print(output)
if self.config.have_patience:
if val_loss < best_validation_loss :
if val_loss < best_validation_loss * improvement_threshold :
self.saver.save(sess, self.config.ckptdir_path + "/model_best.ckpt")
best_validation_loss = val_loss
best_step = self.epoch_count
else:
if patience < 1:
self.saver.restore(sess, self.config.ckptdir_path + "/model_best.ckpt")
if learning_rate <= 0.00001 :
print("=> Breaking by Patience Method")
break
else :
learning_rate /= 10
patience = self.config.patience
print("\033[91m=> Learning rate dropped to {}\033[0m".format(learning_rate))
else :
patience -= 1
self.epoch_count += 1
print("=> Best epoch : {}".format(best_step))
if self.config.debug == True:
sys.exit()
test_loss = self.run_eval(sess, "test", summarizer['test'], tr_step)
test_metrics, test_accuracy, p_k = self.get_metrics(sess, "test")
returnDict = {"test_loss" : test_loss, "test_accuracy" : test_accuracy, 'test_metrics' : test_metrics, "test_pak" : p_k}
if self.config.debug == False:
returnDict["train"] = best_validation_loss
self.saver.save(sess, self.config.ckptdir_path)
return returnDict
class Gen():
def __init__(self, value: list, max_count: int, max_lauers: int,
min_lauers: int, k: int) -> None:
self.max_count = max_count
self.max_lauers = max_lauers
self.min_lauers = min_lauers
self.k = k
if value:
self.value = value
else:
# Если не передана стурктура нейронной сети, генерируем ее
self.value = [10]
self.value += [
randint(1, self.max_count)
for _ in range(randint(min_lauers - 2, max_lauers - 2))
]
self.value += [1]
# Создание и обучение нейронной сети
def create(self, epochs: int) -> None:
dataset = pd.read_csv('housepricedata.csv').values
X = dataset[:, 0:10]
Y = dataset[:, 10]
min_max_scaler = preprocessing.MinMaxScaler()
X_scale = min_max_scaler.fit_transform(X)
X_train, X_val_and_test, Y_train, Y_val_and_test = train_test_split(
X_scale, Y, test_size=0.3)
X_val, self.X_test, Y_val, self.Y_test = train_test_split(
X_val_and_test, Y_val_and_test, test_size=0.5)
self.network = Network(self.value)
self.network.compile()
self.network.fit(X_train, Y_train, self.value[0], epochs,
(X_val, Y_val))
# Расчет значения для отбора геномов
def check(self) -> float:
return self.network.accuracy() * self.k / sum(self.value)
# Мутация гена
def mutation(self) -> None:
operator = randint(0, 2)
if operator == 0 and len(self.value) > 2: # Изменение занчения
self.value[randint(1,
len(self.value) - 2)] = randint(
1, self.max_count)
elif operator == 1 and len(
self.value) < self.max_lauers: # Добавление слоя
self.value.insert(randint(1,
len(self.value) - 1),
randint(1, self.max_count))
elif operator == 2 and len(
self.value) > self.min_lauers: # Удаление слоя
del self.value[randint(1, len(self.value) - 2)]
# Получение точности сети
def accuracy(self) -> float:
return self.network.accuracy()
# Получение ошибки сети
def mse(self) -> float:
return self.network.mse(self.X_test, self.Y_test)
def main():
tries = 0
plays = 0
db_good = []
db_bad = []
minesweeper = Minesweeper()
# stores XY of mouse events
mouse_x = 0
mouse_y = 0
while True:
has_game_ended = False
game_over = False
minesweeper.new_game()
tries += 1
print('Jogo', tries)
# For random training
chosen_squares = []
chosen_squares_length = 1
# Main game loop
while not has_game_ended:
# Initialize variables
mouse_clicked = False
safe_squares = []
flagged_squares = []
new_position = False # For random training
# Draw screen
minesweeper.draw_field()
# Get Fred's AI input
if AI_TYPE == 'FRED':
info = minesweeper.available_info()
safe_squares, flagged_squares = minesweeper.get_AI_input(info)
# Get random position
if AI_TYPE == 'RANDOM':
while not new_position:
choice = [[random.choice(range(FIELDWIDTH)),random.choice(range(FIELDHEIGHT))]]
if minesweeper.revealed_boxes[choice[0][0]][choice[0][1]] == False:
new_position = True
safe_squares = choice
# Get player input
for event in pygame.event.get():
if event.type == QUIT or (event.type == KEYDOWN and (event.key == K_ESCAPE or event.key == K_q)):
minesweeper.terminate()
elif event.type == MOUSEMOTION:
mouse_x, mouse_y = event.pos
elif event.type == MOUSEBUTTONDOWN:
if event.button == LEFT_CLICK:
mouse_x, mouse_y = event.pos
mouse_clicked = True
box_x, box_y = minesweeper.get_box_at_pixel(mouse_x, mouse_y)
if box_x is not None and box_y is not None:
safe_squares = [(box_x, box_y)]
if event.button == RIGHT_CLICK:
mouse_x, mouse_y = event.pos
box_x, box_y = minesweeper.get_box_at_pixel(mouse_x, mouse_y)
if box_x is not None and box_y is not None:
flagged_squares = [(box_x, box_y)]
# Keeps track of chosen squares
if len(safe_squares) > 0:
chosen_squares.append(safe_squares)
# Checks if game is over for AI
if AI_TYPE != 'HUMAN':
if game_over:
has_game_ended = True
# Saves turn
if TRAINING and chosen_squares_length == len(chosen_squares):
turn = minesweeper.save_turn()
# Apply game changes
if not game_over:
for x, y in flagged_squares:
minesweeper.toggle_flag_box(x, y)
for x, y in safe_squares:
minesweeper._RESET_SURF, minesweeper._RESET_RECT = minesweeper.draw_smiley(WINDOWWIDTH/2, 50, 'check.png')
game_over = minesweeper.reveal_box(x, y)
#Add play to DB
if TRAINING and chosen_squares_length == len(chosen_squares) and len(safe_squares) > 0:
chosen_squares_length += 1
turn_chosen_square = minesweeper.save_chosen_square(safe_squares)
if minesweeper.mine_field[safe_squares[0][0]][safe_squares[0][1]] != 'X':
db_good.append((np.asarray(turn),np.asarray(turn_chosen_square)))
#print('GOOD', db_good[-1])
if minesweeper.mine_field[safe_squares[0][0]][safe_squares[0][1]] == 'X':
db_bad.append((np.asarray(turn),np.asarray(turn_chosen_square)))
#print('BAD', db_bad[-1])
plays += 1
if plays > TRAINING_COMPLETE:
db_good = minesweeper.data_treat(db_good)
db_bad = minesweeper.data_treat(db_bad)
print('GOOD',db_good[0])
print('BAD',db_bad[0])
div = int(len(db_good)*.8)
training_data = db_good[:div]
test_data = db_good[div:]
net = Network([FIELDHEIGHT*FIELDWIDTH, 32, 16, FIELDHEIGHT*FIELDWIDTH])
print('Network training started...\n')
net.SGD(training_data, 30, 10, 0.1, lmbda=5.0)
net.save('ANN_test')
print('score = {:.2f} %'.format(net.accuracy(test_data)/100))
minesweeper.terminate()
# Check if reset box is clicked
if minesweeper._RESET_RECT.collidepoint(mouse_x, mouse_y):
minesweeper.highlight_button(minesweeper._RESET_RECT)
if mouse_clicked:
minesweeper.new_game()
has_game_ended = True
# Highlight unrevealed box
box_x, box_y = minesweeper.get_box_at_pixel(mouse_x, mouse_y)
if box_x is not None and box_y is not None and not minesweeper.revealed_boxes[box_x][box_y]:
minesweeper.highlight_box(box_x, box_y)
# Update screen, wait clock tick
pygame.display.update()
minesweeper.clock.tick(FPS)
from initializers import Xavier
from layers import *
from network import Network
cifar = CIFAR10()
training = cifar.get_named_batches('data_batch_1').subset(10)
net = Network()
net.add_layer(Linear(CIFAR10.input_size, 50, 0, Xavier()))
net.add_layer(ReLU(50))
net.add_layer(Linear(50, CIFAR10.output_size, 0, Xavier()))
net.add_layer(Softmax(CIFAR10.output_size))
Y = net.evaluate(training.images)
print('Cost:', net.cost(training.one_hot_labels, None, Y))
print('Accuracy: {:.2%}'.format(net.accuracy(training.one_hot_labels, None,
Y)))
plt.subplot(1, 3, 1)
plt.imshow(Y)
plt.yticks(range(10), cifar.labels)
plt.xlabel('Image number')
plt.title('Probabilities')
plt.subplot(1, 3, 2)
plt.imshow(cifar.label_encoder.transform(np.argmax(Y, axis=0)).T)
plt.yticks([])
plt.xlabel('Image number')
plt.title('Predicted classes')
plt.subplot(1, 3, 3)
plt.imshow(training.one_hot_labels)
print("")
print("")
print("=========================================")
print("==== Start training")
print("=========================================")
iteration = tuple(range(iters_num))
for i in iteration:
# Calculate gradients.
grad = network.gradient(x_train[i], t_train[i])
# Update parameters.
for key in network.params:
network.params[key] -= learning_rate * grad[key]
if i % iter_per_epoch == 0:
print("iteration: " + str(i), end='\t')
i_list.append(i)
train_acc = network.accuracy(x_train, t_train)
test_acc = network.accuracy(x_test, t_test)
test_acc_list.append(test_acc)
print("Accuracy of training data: {0:.5f}".format(train_acc), end='\t')
print("test data: {0:.5f}".format(test_acc))
print("")
print("")
print("=========================================")
print("==== Report result")
print("=========================================")
def show_image(img):
for i in img:
for j in i:
import pickle
from mnist_loader import load_data_wrapper
from network import Network
if __name__ == '__main__':
# Load weights and biases
with open("temp", "rb") as f:
W, B = pickle.load(f)
net = Network([784, 50, 10])
# Use the weights and biases from the previously trained network
net.W = W
net.B = B
test_data = load_data_wrapper()[2]
accuracy = net.accuracy(test_data)
print("Nework accuracy is: %f%%" % (accuracy * 100))
Dense(size=10, input_len=2048), Softmax())
optimizer = Adam(network.trainables,
learning_rate=lambda n: 0.0001,
beta_1=0.9,
beta_2=0.999)
avg = IncrementalAverage()
for epoch in range(STARTING_EPOCH, STARTING_EPOCH + EPOCHS):
batch = 1
for x, y in make_batch(training_data, training_labels, BATCH_SIZE):
out = network(x)
avg.add(np.sum(VectorCrossEntropy.error(out, y)))
network.backward(VectorCrossEntropy.gradient(out, y), update=True)
if batch % LOG_FREQ == 0:
print(
f"epoch {epoch}/{EPOCHS} | batch {batch} - loss: {avg.get()}")
batch += 1
# Testing
testacc = IncrementalAverage()
# Split the test data into 10 batches in order to fit in RAM.
for testbatch in range(0, 10000, 1000):
testacc.add(
network.accuracy(test_data[testbatch:testbatch + 1000],
test_labels[testbatch:testbatch + 1000]))
accuracy = testacc.get()
print(f"Test accuracy : {accuracy}")
# network.save_weights("path-to-savefile")
print(f"Epoch {epoch} done =================")
avg.reset()
net = Network(model_path, training_dropout=0.5)
# debugging to keep track of accuracy even through checkpoints
graph = []
i = 0
while True:
print ('step {:,}'.format(i), end='\r')
# load up to 10 data points with random distribution of
# true/false data in random proportions (averaging at 50/50 True/False)
batch = rearrange_batch(training_data.load(10))
# debugging accuracy printout
if i % print_time == 0:
batch = rearrange_batch(training_data.load(1000000))
train_accuracy = net.accuracy(batch[0], batch[1], batch[2])
print('step {:,}, training accuracy {:,.6f}'.format(i, train_accuracy))
graph.append(train_accuracy)
# save graph
if i and i % save_time == 0:
print ('step {:,} saving (average = {:,.5f})'.format(i, sum(graph)/len(graph)))
net.save()
with open(model_path + '.csv', 'a') as f:
# Save the accuracies so that we can check them later on
f.write('\n'.join(str(s) for s in graph) + '\n')
graph = []
# training
net.train(batch[0], batch[1], batch[2])
i += 1
class Model(object):
def __init__(self, config):
self.epoch_count = 0
self.config = config
self.data = DataSet(config)
self.add_placeholders()
self.summarizer = tf.summary
self.net = Network(config, self.summarizer)
self.optimizer = self.config.solver.optimizer
self.y_pred = self.net.prediction(self.x, self.keep_prob)
self.loss = self.net.loss_function(self.x, self.y, self.keep_prob)
self.accuracy = self.net.accuracy(self.y_pred, self.y)
self.summarizer.scalar("accuracy", self.accuracy)
self.summarizer.scalar("loss", self.loss)
self.train = self.net.train_step(self.loss)
self.B = self.net.B
self.A = self.net.A
self.n_epoch_to_decay = list(range(800, 20000, 1000))[::-1]
self.next_epoch_to_decay = self.n_epoch_to_decay.pop()
self.saver = tf.train.Saver()
self.init = tf.global_variables_initializer()
self.local_init = tf.local_variables_initializer()
self.kf = KFold(n_splits=10, random_state=0, shuffle=True)
def add_placeholders(self):
self.x = tf.placeholder(tf.float32, shape=[None, self.config.features_dim])
self.y = tf.placeholder(tf.float32, shape=[None, self.config.labels_dim])
self.keep_prob = tf.placeholder(tf.float32)
def train_epoch(self, sess, summarizer):
merged_summary = self.summarizer.merge_all()
err, accuracy = list(), list()
X, Y = self.data.get_train()
for train, val in self.kf.split(X, y=Y):
feed_dict = {self.x: X[train], self.y: Y[train], self.keep_prob: self.config.solver.dropout}
# attention!
summ, _, loss_, accuracy_ = sess.run([merged_summary, self.train,
self.loss, self.accuracy], feed_dict=feed_dict)
summarizer.add_summary(summ)
err.append(loss_)
accuracy.append(accuracy_)
return np.mean(err), np.mean(accuracy)
def do_eval(self, sess, data):
if data == "validation":
err, accuracy = list(), list()
X, Y = self.data.get_validation()
for train, val in self.kf.split(X, y=Y):
feed_dict = {self.x: X[val], self.y: Y[val], self.keep_prob: 1}
loss_, Y_pred, accuracy_ = sess.run([self.loss, self.y_pred, self.accuracy], feed_dict=feed_dict)
metrics = evaluate(predictions=Y_pred, labels=Y[val])
err.append(loss_)
accuracy.append(accuracy_)
return np.mean(err), np.mean(accuracy), metrics
if data == "test":
X, Y = self.data.get_test()
feed_dict = {self.x: X, self.y: Y, self.keep_prob: 1}
loss_, Y_pred, accuracy_ = sess.run([self.loss, self.y_pred, self.accuracy], feed_dict=feed_dict)
metrics = evaluate(predictions=Y_pred, labels=Y)
return loss_, accuracy_, metrics
def fit(self, sess, summarizer):
sess.run(self.init)
sess.run(self.local_init)
max_epochs = self.config.max_epochs
self.epoch_count = 0
max_micro_f1 = 0
max_macro_f1 = 0
while self.epoch_count < max_epochs:
if self.config.load:
break
loss_train, accuracy_train = self.train_epoch(sess, summarizer['train'])
loss_val, accuracy_val, metrics_val = self.do_eval(sess, "validation")
if self.epoch_count == self.next_epoch_to_decay:
if len(self.n_epoch_to_decay) == 0:
self.next_epoch_to_decay = -1
else:
self.next_epoch_to_decay = self.n_epoch_to_decay.pop()
self.config.learning_rate *= self.config.lr_decay_factor
print('Decaying learning rate ...')
print(self.config.learning_rate)
if max_micro_f1 < metrics_val['micro_f1'] and max_macro_f1 < metrics_val['macro_f1']:
print(self.config.ckptdir_path)
print("cur_max_Mi-F1 = %g, cur_max_Ma-F1 = %g, cur_epoch = %g." % (
metrics_val['micro_f1'], metrics_val['macro_f1'], self.epoch_count))
self.saver.save(sess, self.config.ckptdir_path + "model.ckpt")
max_micro_f1 = max(max_micro_f1, metrics_val['micro_f1'])
max_macro_f1 = max(max_macro_f1, metrics_val['macro_f1'])
if self.epoch_count % 5 == 0:
print("After %d training epoch(s), Training : Loss = %g, Validation : Loss = %g." % (
self.epoch_count, loss_train, loss_val))
print("train_accuracy = %g, val_accuracy = %g." % (accuracy_train, accuracy_val))
print("Micro-F1 = %g, Macro-F1 = %g." % (metrics_val['micro_f1'], metrics_val['macro_f1']))
self.epoch_count += 1
returnDict = {"train_loss": loss_train, "val_loss": loss_val, "train_accuracy": accuracy_train,
"val_accuracy": accuracy_val}
return returnDict
def add_summaries(self, sess):
if self.config.load or self.config.debug:
path_ = os.path.join("../results/tensorboard" + self.config.dataset_name)
else:
path_ = os.path.join("../bin/results/tensorboard" + self.config.dataset_name)
summary_writer_train = tf.summary.FileWriter(path_ + "/train", sess.graph)
summary_writer_val = tf.summary.FileWriter(path_ + "/val", sess.graph)
summary_writer_test = tf.summary.FileWriter(path_ + "/test", sess.graph)
summary_writers = {'train': summary_writer_train, 'val': summary_writer_val, 'test': summary_writer_test}
return summary_writers
