def run():
# Initialize size of network with
inode = 3
hnode = 4
onode = 2
n = NeuralNetwork(inode, hnode, onode)
entry = [1.0, 0.5, -1.5]
result = n.query(entry)
print("Init:")
print(n.w_input_hidden)
target = list(result.T[0])
n.train(entry, target)
print("")
print("Target:")
print(n.w_input_hidden)
target[0] += 1
n.train(entry, target)
print("")
print("Error:")
print(n.w_input_hidden)
def __init__(self, host_str, model_path):
# load trained neural network
self.nn = NeuralNetwork()
self.nn.load_model(model_path)
self.car_ctrl = sock_car_control()
def run(data_obj, training_size):
data_obj.split_dataset(training_size)
data_obj.preprocess()
#-------------------------------- Autoencoder model
ae_model = AutoEncoder(data_obj.x_train_scaled.shape[1],
training_size, data_obj.name)
ae_model.train(data_obj.x_train_scaled, data_obj.x_val_scaled)
#-------------------------------- Encoded representation
x_train_encoded, x_val_encoded, x_test_encoded = ae_model.encoded_data(
data_obj.x_train_scaled, data_obj.x_val_scaled, data_obj.x_test_scaled)
#-------------------------------- Neural Network model
nn_model = NeuralNetwork(
data_obj.x_train_scaled.shape[1], data_obj.y_train.shape[1],
training_size, data_obj.name)
nn_model.train(
x_train_encoded, data_obj.y_train, x_val_encoded, data_obj.y_val)
nn_model.evaluate(x_test_encoded, data_obj.y_test)
#-------------------------------- reset data from memory
data_obj.reset_scalar()
return nn_model.result()
def __init__(self, host, port, _host, _port, model_path):
# load trained neural network
self.nn = NeuralNetwork(model_path)
self.cap = cv2.VideoCapture(0)
self.rc_car = RCControl(_h, _p)
def use_QII_with_model(X_train):
model = NeuralNetwork()
model.create_graph()
hl_1, hl_2, hl_3, prediction = model.predict(X_train)
numb_features = 4 # hidden layer 3 has 4 units
feature_list = [str(i) for i in range(1, numb_features + 1)]
influencer = QII(hl_3, feature_list)
# value of one data points we want to observe
data_point = hl_3[0]
with tf.Session(graph=model.graph) as sess:
saver = tf.train.Saver()
saver.restore(sess, model.CHECK_POINT)
print('Correct value with Exhaustive Computation')
banzhaf_val = influencer.banzhaf(data_point,
model.predict_hl_3,
sess,
is_exhaustive=True)
print(banzhaf_val)
influence_scores = np.array([banzhaf_val[i] for i in feature_list])
influence_scores = np.expand_dims(influence_scores, axis=0)
weight_name_list = ['w1', 'w2', 'w3']
influence_scores = influencer.mapback(model, X_train,
influence_scores,
weight_name_list)
influencer.plot_influence_score(influence_scores, feature_list)
def test_run(self):
# Test correctness of run method
network = NeuralNetwork(3, 2, 1, 0.5)
network.weights_input_to_hidden = test_w_i_h.copy()
network.weights_hidden_to_output = test_w_h_o.copy()
self.assertTrue(np.allclose(network.run(inputs), 0.09998924))
def test_case_2():
import time
dataset = IRISDataset()
X_test, X_train, y_test, y_train = dataset.split_to_train_test(seed=10)
model = NeuralNetwork()
model.create_graph()
hl_1, hl_2, hl_3, prediction = model.predict(X_train)
numb_features = 8
feature_list = [str(i) for i in range(1, numb_features + 1)]
influencer = QII(hl_1, feature_list)
data_point = hl_1[0]
with tf.Session(graph=model.graph) as sess:
saver = tf.train.Saver()
saver.restore(sess, model.CHECK_POINT)
start_time = time.time()
print('Approximation with Random Sampling')
shapley_val = influencer.banzhaf(data_point,
model.predict_hl_1,
sess,
is_exhaustive=False)
print(shapley_val)
print("--- %s seconds ---" % (time.time() - start_time))
start_time = time.time()
print('Correct value with Exhaustive Computation')
shapley_val = influencer.banzhaf(data_point,
model.predict_hl_1,
sess,
is_exhaustive=True)
print(shapley_val)
print("--- %s seconds ---" % (time.time() - start_time))
def main():
"""Main entry to the application
"""
args = get_console_args()
# Define the network's architecture
architecture = JsonProcessor.load(args.architecture_file)
# Initialize the Neural Network
nn = NeuralNetwork(**architecture)
# Set its parameters
nn.set_parameters()
# Loads the data
nn.load_data(args.training_data)
# Train the neural network
nn.train(args.epochs)
# Get its predictions
predictions = nn.predict(JsonProcessor.load(args.testing_data)['input'])
# Save its predictions
JsonProcessor.beautify(args.output_file, {'output': predictions.tolist()})
def use_QII_with_model(X_train):
model = NeuralNetwork()
model.create_graph()
hl_1, hl_2, hl_3, prediction = model.predict(X_train)
numb_features = 4 # hidden layer 3 has 4 units
feature_list = [str(i) for i in range(1, numb_features + 1)]
influencer = QII(hl_3, feature_list)
# value of one data points we want to observe
data_point = hl_3[0]
assignment_list = ['normalized', 'ranking', 'nochange']
propagate_list = [
'transpose_mul', 'inverse_mul', 'softmax_dist', 'shift_dist',
'shift_softmax_dist'
]
with tf.Session(graph=model.graph) as sess:
saver = tf.train.Saver()
saver.restore(sess, model.CHECK_POINT)
for assignment in assignment_list:
for propagate in propagate_list:
banzhaf_val = influencer.banzhaf(data_point,
model.predict_hl_3,
sess,
is_exhaustive=False)
influence_scores = np.array(
[banzhaf_val[i] for i in feature_list])
influence_scores = np.expand_dims(influence_scores, axis=0)
weight_name_list = ['w1', 'w2', 'w3']
influence_scores = influencer.mapback(
model, X_train, influence_scores, weight_name_list,
assignment, propagate)
influencer.plot_influence_score(influence_scores,
feature_list)
def __init__(self, host, port, model_path):
self.server_socket = socket.socket()
self.server_socket.bind((host, port))
self.server_socket.listen(0)
# accept a single connection
self.connection = self.server_socket.accept()[0].makefile('rb')
self.client_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
self.client_socket.connect(('192.168.0.115', 1234)) #pi
# load trained neural network
self.nn = NeuralNetwork()
self.nn.load_modelKeras("model_test.h5")
self.h1 = 5.5 #stop sign - measure manually
#self.stop_cascade = cv2.CascadeClassifier(cv2.data.haarcascades +"C:/Users/My\ PC/Anaconda3/envs/tf/Lib/site-packages/cv2/data/stop_sign.xml")
#self.stop_cascade = cv2.CascadeClassifier(cv2.data.haarcascades +"D:/Downloads/self-driving-car2/neural networks/stop_sign.xml")
self.stop_cascade = cv2.CascadeClassifier(cv2.data.haarcascades +
"stop_sign.xml")
self.d_stop_light_thresh = 70
self.d_stop_sign = self.d_stop_light_thresh
self.stop_start = 0 # start time when stop at the stop sign
self.stop_finish = 0
self.stop_time = 0
self.drive_time_after_stop = 0
self.alpha = 8.0 * math.pi / 180 # degree measured manually
self.v0 = 119.865631204 # from camera matrix
self.ay = 332.262498472 # from camera matrix
def main():
# load dataset
adata = load_data('./data/')
train_adata, test_adata = train_test(adata, 0.8)
dataset = KidneyDataset('./data/')
train_set, test_set = random_split(dataset, [math.floor(len(dataset) * 0.8), len(dataset) - math.floor(len(dataset) * 0.8)])
# not sure about batch sizes, need to figure this out more
train_dl = DataLoader(train_set, batch_size=64)
test_dl = DataLoader(test_set, batch_size=64)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f'Using {device} device')
model = NeuralNetwork()
model = model.to(device)
wandb.watch(model)
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
for i in range(4):
print(f"Epoch {i+1}\n-------------------------------")
train(train_dl, model, loss_fn, optimizer)
test(test_dl, model)
def joint_actuate(clientID, Body, primitive, frame):
model = NeuralNetwork()
model.load_state_dict(torch.load('C:\\Users\\lenovo\\Desktop\\AI-Project-Portfolio\\Amendments\\new_angle_mapping\\weights_2.pth'))
model.eval()
angle = model(torch.from_numpy(primitive[frame].reshape(1, -1)).float())
HP, HY, LSP, LSR, LER, LEY, RSP, RSR, RER, REY, LHP, LHR, LHYP, LKP, RHP, RHR, RHYP, RKP, LAP, LAR, RAP, RAR = angle.detach().numpy().reshape(-1)
print(LSP, LSR, LER, LEY)
def control(joint, angel):
sim.simxSetJointTargetPosition(clientID, Body[joint], angel, sim.simx_opmode_oneshot)
sim.simxPauseCommunication(clientID,1)
control('HeadPitch', HP)
control('HeadYaw', HY)
control('LShoulderPitch', 1*LSP)
control('LShoulderRoll', 1*LSR)
control('LElbowRoll', 1*LER)
control('LElbowYaw', 1*LEY)
control('RShoulderPitch', RSP)
control('RShoulderRoll', RSR)
control('RElbowRoll', RER)
control('RElbowYaw', REY)
control('LHipPitch', LHP)
control('LHipRoll', LHR)
control('LHipYawPitch', LHYP)
control('LKneePitch', LKP)
control('RHipPitch', RHP)
control('RHipRoll', RHR)
control('RHipYawPitch', RHYP)
control('RKneePitch', RKP)
control('LAnklePitch', LAP)
control('LAnkleRoll', LAR)
control('RAnklePitch', RAP)
control('RAnkleRoll', RAR)
sim.simxPauseCommunication(clientID,0)
#time.sleep(.04)
def worker(args, STD_NOISE, STATE_DIM, ACTION_DIM, params_queue, output_queue,
elite_queue):
# Function execute by each worker: get the agent' NN, sample noise and evaluate the agent adding the noise. Then return the seed and the rewards to the central unit
env = gym.make(args.env)
# env = CyclicMDP()
actor = NeuralNetwork(STATE_DIM, ACTION_DIM.shape[0], args.hidden_size)
while True:
# get the new actor's params
act_params = params_queue.get()
if act_params != None:
# load the actor params
actor.load_state_dict(act_params)
# get a random seed
seed = np.random.randint(1e6)
# set the new seed
np.random.seed(seed)
noise = sample_noise(actor)
pos_rew = evaluate_noisy_net(STD_NOISE, noise, actor, env,
elite_queue)
# Mirrored sampling
neg_rew = evaluate_noisy_net(STD_NOISE, -noise, actor, env,
elite_queue)
output_queue.put([[pos_rew, neg_rew], seed])
else:
break
def train_model(X_test, X_train, y_test, y_train):
print_steps = 50
epoch = 100
model = NeuralNetwork()
model.initialize_variables()
model.create_graph()
model.train(X_train, y_train, print_steps, epoch)
model.test(X_test, y_test)
return model
class VideoStreamHandler(socketserver.StreamRequestHandler):
# load trained neural network
model = NeuralNetwork()
model.modified_model()
model.load_model()
# # hard coded thresholds for stopping, sensor 30cm, other two 25cm
d_sensor_thresh = 15
def handle(self):
global prediction
global sensor_data
stream_bytes = b' '
try:
# stream video frames one by one
while True:
stream_bytes += self.rfile.read(1024)
first = stream_bytes.find(b'\xff\xd8')
last = stream_bytes.find(b'\xff\xd9')
if first != -1 and last != -1:
jpg = stream_bytes[first:last + 2]
stream_bytes = stream_bytes[last + 2:]
gray = cv2.imdecode(np.frombuffer(jpg, dtype=np.uint8),
cv2.IMREAD_GRAYSCALE)
image = cv2.imdecode(np.frombuffer(jpg, dtype=np.uint8),
cv2.IMREAD_COLOR)
# lower half of the image
height, width = gray.shape
roi = gray[int(height / 1.25):height, :]
cv2.imshow('image', image)
# reshape image
gray = (np.expand_dims(roi, 0))
gray = gray[:, :, :, np.newaxis]
# neural network makes prediction
prediction = self.model.predict(gray)
# stop conditions
if sensor_data and int(sensor_data) < self.d_sensor_thresh:
print("Dur, Nesne Algılandı...!\n")
prediction = 4
sensor_data = None
if cv2.waitKey(1) & 0xFF == ord('q'):
print("car stopped")
break
finally:
cv2.destroyAllWindows()
sys.exit()
def __init__(self, client, steer):
self.client = client
self.steer = steer
self.stopline = StopLine.Stop()
self.dect = Object.Object_Detection(self.steer)
# model create
self.model = NeuralNetwork()
self.model.load_model(path = 'model_data/video_model_1.h5')
def crossover(args, STATE_DIM, ACTION_DIM, gene1, gene2):
actor_1 = NeuralNetwork(STATE_DIM, ACTION_DIM.shape[0], args.hidden_size)
actor_1.load_state_dict(gene1)
actor_2 = NeuralNetwork(STATE_DIM, ACTION_DIM.shape[0], args.hidden_size)
actor_2.load_state_dict(gene2)
for param1, param2 in zip(actor_1.parameters(), actor_2.parameters()):
# References to the variable tensors
W1 = param1.data
W2 = param2.data
if len(W1.shape) == 2: #Weights no bias
num_variables = W1.shape[0]
# Crossover opertation [Indexed by row]
num_cross_overs = random.randrange(
num_variables * 2) # Lower bounded on full swaps
for i in range(num_cross_overs):
receiver_choice = random.random(
) # Choose which gene to receive the perturbation
if receiver_choice < 0.5:
ind_cr = random.randrange(W1.shape[0]) #
W1[ind_cr, :] = W2[ind_cr, :]
else:
ind_cr = random.randrange(W1.shape[0]) #
W2[ind_cr, :] = W1[ind_cr, :]
elif len(W1.shape) == 1: #Bias
num_variables = W1.shape[0]
# Crossover opertation [Indexed by row]
num_cross_overs = random.randrange(
num_variables) # Lower bounded on full swaps
for i in range(num_cross_overs):
receiver_choice = random.random(
) # Choose which gene to receive the perturbation
if receiver_choice < 0.5:
ind_cr = random.randrange(W1.shape[0]) #
W1[ind_cr] = W2[ind_cr]
else:
ind_cr = random.randrange(W1.shape[0]) #
W2[ind_cr] = W1[ind_cr]
def train(args, data_info):
n_user = data_info[0]
n_item = data_info[1]
train_data = data_info[2]
val_data = data_info[3]
test_data = data_info[4]
model = NeuralNetwork(args, n_user, n_item)
train_rmse_ls = []
val_rmse_ls = []
test_rmse_ls = []
epoch = []
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for step in range(args.n_epochs):
np.random.shuffle(train_data)
start = random.randint(0, train_data.shape[0] - args.batch_size)
_ = model.train(sess,
feed_dict=get_feed_dict(train_data, model, start,
start + args.batch_size))
#start = 0
#while start < len(train_data):
# model.train(sess, feed_dict=get_feed_dict(train_data, model, start, start + args.batch_size))
# start += args.batch_size
train_loss, train_rmse = model.validation(sess,
feed_dict=get_feed_dict(
train_data, model, 0,
len(train_data)))
val_loss, val_rmse = model.validation(sess,
feed_dict=get_feed_dict(
val_data, model, 0,
len(val_data)))
test_loss, test_rmse = model.validation(sess,
feed_dict=get_feed_dict(
test_data, model, 0,
len(test_data)))
train_rmse_ls.append(train_rmse)
val_rmse_ls.append(val_rmse)
test_rmse_ls.append(test_rmse)
epoch.append(step)
print(
'epoch %d train loss: %.4f rmse: %.4f val loss: %.4f rmse: %.4f test loss: %.4f rmse: %.4f'
% (step, train_loss, train_rmse, val_loss, val_rmse, test_loss,
test_rmse))
return train_rmse_ls, val_rmse_ls, test_rmse_ls, epoch
def test_train(self):
# Test that weights are updated correctly on training
network = NeuralNetwork(3, 2, 1, 0.5)
network.weights_input_to_hidden = test_w_i_h.copy()
network.weights_hidden_to_output = test_w_h_o.copy()
network.train(inputs, targets)
self.assertTrue(np.allclose(network.weights_hidden_to_output,
np.array([[ 0.37275328],
[-0.03172939]])))
self.assertTrue(np.allclose(network.weights_input_to_hidden,
np.array([[ 0.10562014, -0.20185996],
[0.39775194, 0.50074398],
[-0.29887597, 0.19962801]])))
def __init__(self, host, port, serial_port, model_path):
self.server_socket = socket.socket()
self.server_socket.bind((host, port))
self.server_socket.listen(0)
# accept a single connection
self.connection = self.server_socket.accept()[0].makefile('rb')
# load trained neural network
self.nn = NeuralNetwork()
self.nn.load_model(model_path)
self.rc_car = RCControl(serial_port)
def __init__(self, hidden_sizes, batch_size, learning_rate, activator):
"""The entry of project.
Load data and build Neural Network model.
"""
self.hidden_sizes = hidden_sizes
self.batch_size = batch_size
self.learning_rate = learning_rate
self.activator = activator
self.data = Data(SHELL_ARGS.prefix)
self.model = NeuralNetwork(self.data.input_size(), self.hidden_sizes,
self.data.output_size(), self.batch_size,
self.activator)
def main():
# Load dataset.
data = load_data("MNISTdata.hdf5")
train_data, test_data = data[:2], data[2:]
nn = NeuralNetwork(train_data[0].shape[1], hidden_units=100)
# Train.
print("training Network...")
nn.train(train_data, learning_rate1=0.1, epochs=20)
# Test
print("testing Network...")
nn.test(test_data)
def __init__(self, client, steer):
self.client = client
self.steer = steer
self.stopline = StopLine.Stop()
self.objDetect = Object.Object_Detection(
self.steer) # class - load object detection model
# model create
print("Load VGGNet")
self.model = NeuralNetwork()
self.model.load_model(
path='../model_data/posicar_binary_v11.h5') # self-driving model
print("Complete")
def __init__(self, model_path):
print('initing...')
# load trained neural network
if model_num == 1:
import model_b as model
self.model = model
elif model_num == 0:
load_model_start = time.time()
from model import NeuralNetwork
self.model = NeuralNetwork(model_path)
load_model_end = time.time()
print('loading model costs {:02f} second(s)'.format(load_model_end - load_model_start))
self.cap = cv2.VideoCapture(0)
self.servo = PyServo.Servo(SerialID,Baudrate)
def main():
"""High level pipeline."""
# Load dataset.
data = load_data("data/MNISTdata.hdf5")
train_data, test_data = data[:2], data[2:]
# Init model.
nn = NeuralNetwork(train_data[0].shape[1], hidden_units=100)
# Train.
print("training model...")
nn.train(train_data, learning_rate=0.1, epochs=25)
# Test
print("testing model...")
nn.test(test_data)
def __init__(self,
layer_sizes,
env_name,
invert_reward=True,
amount_threads=None,
cpu_only=True,
number_repeats=5,
seed=8,
reward_offset=0):
self.reward_offset = reward_offset
self.invert_reward = invert_reward
self.amount_threads = multiprocessing.cpu_count(
) if amount_threads == None else amount_threads
self.envs = [gym.make(env_name) for _ in range(self.amount_threads)]
for env in self.envs:
env.seed(seed)
self.env_name = env_name
input_size = self.envs[0].observation_space.shape[0]
if isinstance(self.envs[0].action_space, spaces.Discrete):
print(env_name, "has a discrete action space")
output_activation = softmax
output_size = self.envs[0].action_space.n
self.discrete_action = True
output_activation = softmax
else:
print(env_name, "has a non-discrete action space")
output_activation = None
output_size = self.envs[0].action_space.shape[0]
self.discrete_action = False
output_activation = np.tanh
self.models = [
NeuralNetwork(input_size=input_size,
hidden_sizes=layer_sizes,
output_size=output_size,
output_activation=output_activation,
discrete_action=self.discrete_action)
for _ in range(self.amount_threads)
]
self.solution_size = len(self.models[0].get_weights())
self.number_repeats = number_repeats
self.interrupt = False
def create(path, population_size=50, sigma=0.3, learning_rate=0.1, input_shape=(75, 80), output_dim=4):
if os.path.exists(path):
print("Training already exists! Did you maybe wanna continue this training?")
exit()
model = NeuralNetwork(input_shape, output_dim)
training = Training(path, population_size=population_size, sigma=sigma, learning_rate=learning_rate, model=model)
os.makedirs(path)
with open(f"{path}/training_info.txt", "w+") as info_file:
info_file.write(f"{training.iterration}\n")
info_file.write(f"{training.population_size}\n")
info_file.write(f"{training.sigma}\n")
info_file.write(f"{training.learning_rate}\n")
info_file.write(f"{training.best_main}\n")
info_file.write(f"{training.best_all}")
training.model.save(f"{path}/model.h5")
with open(f"{path}/stats.csv", "w+") as _:
pass
return training
def use_QII_with_model(X_train):
model = NeuralNetwork()
model.create_graph()
# get input data for each layer
hl_1, hl_2, hl_3, prediction = model.predict(X_train)
# PREDEFINE PARAMETERS
# Number of features available for sampling, we want to observe
# input layer, hence number of units is 4
numb_features = 4
# initializes QII class with name for each features, here we use index only
feature_list = [str(i) for i in range(1, numb_features + 1)]
influencer = QII(X_train, feature_list)
# value of one data points we want to observe
data_point = X_train[0]
with tf.Session(graph=model.graph) as sess:
saver = tf.train.Saver()
saver.restore(sess, model.CHECK_POINT)
# user can choose to calculate shapley by random sampling of permutation
# or exhaustively check each of them with is_exhaustive
# since 4! = 24 is not a big number, we can try both here
print('Approximation with Random Sampling')
shapley_val = influencer.shapley(data_point,
model.predict_input,
sess,
is_exhaustive=False)
print(shapley_val)
print('Correct value with Exhausive Computation')
shapley_val = influencer.shapley(data_point,
model.predict_input,
sess,
is_exhaustive=True)
print(shapley_val)
def use_QII_with_model(X_train):
model = NeuralNetwork()
model.create_graph()
hl_1, hl_2, hl_3, prediction = model.predict(X_train)
numb_features = 4 # hidden layer 3 has 4 units
feature_list = [str(i) for i in range(1, numb_features + 1)]
influencer = QII(X_train, feature_list)
# value of one data points we want to observe
data_point = X_train[101]
with tf.Session(graph=model.graph) as sess:
saver = tf.train.Saver()
saver.restore(sess, model.CHECK_POINT)
observed_feature_list = [0, 1, 3]
banzhaf_val = influencer.banzhaf(data_point,
model.predict_input,
sess,
observed_feature_list,
is_exhaustive=True)
influence_scores = np.array([banzhaf_val[i] for i in feature_list])
influence_scores = np.expand_dims(influence_scores, axis=0)
influencer.plot_influence_score(influence_scores, feature_list)
def main():
# ----DATA PREPARATION---- #
x, y = make_moons(1000, noise=0.20)
plt.scatter(x[:, 0], x[:, 1], s=10, c=y, cmap=plt.cm.Spectral)
np.random.seed(0)
np.random.shuffle(y)
np.random.seed(0)
np.random.shuffle(x)
# ----MODEL TRAINING---- #
settings = {
'layer_sizes': [x.shape[1], 3, len(np.unique(y))],
'epochs': 100,
'alpha': 0.01,
'lmbda': 0.001
}
nn = NeuralNetwork(**settings)
nn.train(x, y)
