def train():
dataset = "cats"
images = loaddata(dataset)
g, d = loadmodel(dataset)
LR = 0.0002 # initial learning rate
B1 = 0.5 # momentum term
opt = Adam(lr=LR, beta_1=B1)
g.summary()
d.summary()
d.trainable = True
d.compile(loss='binary_crossentropy', optimizer=opt)
g.compile(loss='binary_crossentropy', optimizer=opt)
d.trainable = False
d_on_g = creategans(d, g)
d_on_g.compile(loss='binary_crossentropy', optimizer=opt)
#images=getdata()
#mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
i1, i2, i3, i4 = images.shape
print(i1)
epochs = 40
batch_size = 128 // 2
k = 0
filename = "imscats/" + str(dataset) + "_n_epochs_" + str(epochs) + ".h5"
for i in range(epochs):
g.save(filename)
for j in range(i1 // batch_size):
noise = getnoise(batch_size * 2)
noise_images = g.predict(noise)
d.train_on_batch(images[j * batch_size:(j + 1) * batch_size],
np.ones([batch_size, 1]))
ld = d.train_on_batch(noise_images[0:batch_size],
np.zeros([batch_size, 1]))
if (j % 10 == 0):
print("Epoch: ", i, " D Loss: ", ld)
#d.trainable=False
#d.compile(loss='binary_crossentropy',optimizer=optadam,metrics=['accuracy'])
lg = d_on_g.train_on_batch(noise, np.ones([batch_size * 2, 1]))
#d.trainable=True
#d.compile(loss='binary_crossentropy',optimizer=optadam,metrics=['accuracy'])
if (j % 10 == 0):
print("Epoch: ", i, " G Loss: ", lg)
if (j % 100 == 0):
showim(g, k, noise_test, [i1, i2, i3, i4])
k = k + 1
def train():
"""
Loads datasets {'mnist', 'flowers', 'flowers128',
'cifar10', 'cats'} and their corresponding model
and trains the GANs using 'epochs' epochs.
"""
dataset="mnist"
images=loaddata(dataset)
g,d=loadmodel(dataset)
opt = Adam(lr=0.0002,beta_1=0.5)
g.summary()
d.summary()
d.trainable=True
d.compile(loss='binary_crossentropy', optimizer=opt)
g.compile(loss='binary_crossentropy', optimizer=opt)
d.trainable=False
gansmodel=creategans(d,g)
gansmodel.compile(loss='binary_crossentropy', optimizer=opt)
i1,i2,i3,i4=images.shape
epochs=40
batch_size=128//2
k=0
filename= "ims8/"+str(dataset)+"_n_epochs_"+str(epochs)+".h5"
for i in range(epochs):
g.save(filename)
for j in range(i1//batch_size):
noise=getnoise(batch_size*2)
noise_images=g.predict(noise)
d.train_on_batch(images[j*batch_size:(j+1)*batch_size],np.ones([batch_size,1]))
ld=d.train_on_batch(noise_images[0:batch_size],np.zeros([batch_size,1]))
if (j%10==0):
print("Epoch: ",i+1," D Loss: ",ld)
lg=gansmodel.train_on_batch(noise,np.ones([batch_size*2,1]))
if (j%10==0):
print("Epoch: ",i+1," G Loss: ", lg)
if (j%100==0):
showim(g,k,[i1,i2,i3,i4])
k=k+1
def lambda_check(layers, X_train, Y_train, X_cv, Y_cv):
lambdas = [
0, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32, 0.64, 1.28, 2.56, 5.12, 10.24
]
costs = []
for lambda_regularization in lambdas:
Theta = neuralnetwork.randomtheta(layers, X_train.shape[1])
output = neuralnetwork.train(X_train,
Y_train,
Theta,
lambda_regularization,
maxiterations=233)
cost = neuralnetwork.calculatecost(X_cv, Y_cv, output[0], Theta, 0)
costs.append(cost)
print(costs)
print()
print("Min Cost Is:")
print(lambdas[numpy.argmin(costs)])
matplotlib.pyplot.plot(lambdas, costs)
matplotlib.pyplot.show()
if __name__ == '__main__':
stuff = data.loaddata(sys.argv[1])
# learningcurves([63, 63, 26], stuff.X_train, stuff.Y_train, stuff.X_cv, stuff.Y_cv, 0.16)
lambda_check([63, 63, 26], stuff.X_train, stuff.Y_train, stuff.X_cv,
stuff.Y_cv)
plot_fig(trainX, trainY, testXsave, testY, "my_polynomial_reg")
class PolynomialRegression:
def __init__(self, degree=1):
self.degree = degree
self.coef_ = None
self.intercept_ = None
def _create_phi(self, data):
phi = lambda x: [x**i for i in range(self.degree + 1)]
return np.array([phi(i) for i in data]).reshape(len(data), -1)
def fit(self, data, target):
phi_arr = self._create_phi(data)
w = np.dot(np.linalg.inv(phi_arr.T @ phi_arr), (phi_arr.T @ target))
# 先頭に0を追加
self.coef_ = np.insert(w[1:], 0, 0)
self.intercept_ = w[0]
def predict(self, data):
return (data @ self.coef_) + self.intercept_
if __name__ == '__main__':
trainX, trainY = loaddata("polynomial_reg.csv")
# sklearn
sk_polynomial_reg(trainX, trainY)
# original
my_polynomial_reg(trainX, trainY)
def __cal__distance(self, p0, p1):
return np.sum((p0 - p1)**2)
def fit(self, data, target):
# あらかじめ計算できるものがない...
self._train = data
self._target = target
def predict(self, test):
labels = []
for data in test:
# 距離の計算
distances = np.array(
[self.__cal__distance(p, data) for p in self._train])
# 近い順からn_neighbors個集める
indexes = distances.argsort()[:self.n_neighbors]
# n_neighbors個の中から多数決でラベルを決定
label = Counter(self._target[indexes]).most_common(1)[0][0]
labels.append(label)
return labels
if __name__ == '__main__':
trainX, trainY = loaddata()
# sklearn
calcref(trainX, trainY)
# original
my_knn(trainX, trainY)
from utils import parsargs
from cfgs import cfg, updcfg
from utils import Train
from models import RetinaNet, MiniRetina
from data import loaddata
from pprint import pprint
args = parsargs()
cfg = updcfg(args, cfg)
print('The configs of the model:')
pprint(cfg)
print(cfg.detname)
if cfg.detname == 'retinanet':
print('Loading the retinanet model ...')
model = RetinaNet(cfg)
elif cfg.detname == 'miretina':
print('Loading the miniretina model ...')
model = MiniRetina(cfg)
data = loaddata(cfg)
if args.train:
train = Train(cfg)
train(model, data)