def predict(relevance_vectors, X, mean, kernel_choice):
prediction = []
for xi in range(len(X)):
phi_x = 0
for ri in range(len(relevance_vectors)):
if kernel_choice == "gaussian":
phi_x += mean[ri + 1] * (kernel.gaussian(
X[xi], relevance_vectors[ri])) + mean[0]
elif kernel_choice == "linear":
t1 = X[xi]
phi_x += mean[ri + 1] * (kernel.linear_kernel(
X[xi], relevance_vectors[ri])) + mean[0]
elif kernel_choice == "polynomial":
t1 = X[xi]
phi_x += mean[ri + 1] * (kernel.polynomial_kernel(
X[xi], relevance_vectors[ri])) + mean[0]
elif kernel_choice == "linear_spline":
phi_x += mean[ri + 1] * (kernel.linear_spline(
X[xi], relevance_vectors[ri])) + mean[0]
elif kernel_choice == "rbf":
phi_x += mean[ri + 1] * (kernel.rbf(
X[xi], relevance_vectors[ri])) # + mean[0]
phi_x += mean[0]
prediction.append(phi_x)
return prediction
def main():
try:
image = Image.open(sys.argv[1])
except IOError:
print("Could not open the input \nUsage tick_jpg inputfile.")
sys.exit()
r, g, b = image.split()
rr = np.real(np.array(r))
gr = np.real(np.array(g))
br = np.real(np.array(b))
# too big kern = kernel.gaussian(rr, 30.0)
# kern = kernel.gaussian(rr, 20.0)
kern = kernel.gaussian(rr, 10.0)
kern[0, 0] = 0.0
rp = jacobi_step_with_kernel(rr, kern, 5)
gp = jacobi_step_with_kernel(gr, kern, 5)
bp = jacobi_step_with_kernel(br, kern, 5)
rn = Image.fromarray(np.uint8(rescale(rp, 255.0)))
gn = Image.fromarray(np.uint8(rescale(gp, 255.0)))
bn = Image.fromarray(np.uint8(rescale(bp, 255.0)))
inew = Image.merge("RGB", (rn, gn, bn))
inew.save('after.jpg')
ix = ImageChops.subtract(inew, image, 0.1)
ix.save('difference.jpg')
def test_gaussian_2(): sigma = 2 k = 3 filter_kernel = kernel.gaussian(sigma, k) assert np.sum(filter_kernel) == 1
def test_gaussian_1(): sigma = 2 k = 1 filter_kernel = kernel.gaussian(sigma, k) assert filter_kernel.shape[0] == 1 assert filter_kernel.shape[1] == 2 * k + 1
def main(num_samples=10, num_features=2, grid_size=20):
samples = np.matrix(np.random.normal(size=num_samples * num_features)
.reshape(num_samples, num_features))
labels = 2 * (samples.sum(axis=1) > 0) - 1.0
sigma = 2
trainer =SVMTrainer(kernel.gaussian(sigma))
predictor = trainer.train(samples, labels)
plot(predictor, samples, labels, grid_size)
def main():
try:
original = mrcfile.open(sys.argv[1], mode='r')
output = mrcfile.open(sys.argv[2], mode='w+')
except IOError:
print(
"Could not open the input \nUsage tick_mrc inputfile outputfile.")
sys.exit()
# create a kernel
# 2d version
kern = kernel.gaussian(original.data[0, ], 15.)
kern[0, 0] = 0.
# create list of layers.
layers = []
for layer in original.data:
layers.append(np.float32(jacobi_step_with_kernel(layer, kern, 2)))
output.set_data(np.array(layers))
# output.set_data(original.data)
# output.set_data(np.float32( jacobi_step(original.data, 2)))
# I cannot believe there isn't a method for this
# but there isn't. FTW!
output.header.nx = original.header.nx
output.header.ny = original.header.ny
output.header.nz = original.header.nz
# output.header.mode = original.header.mode
output.header.mode = 2 # for float32
output.header.nxstart = original.header.nxstart
output.header.nystart = original.header.nystart
output.header.nzstart = original.header.nzstart
output.header.mx = original.header.mx
output.header.my = original.header.my
output.header.mz = original.header.mz
output.header.cella = original.header.cella
output.header.cellb = original.header.cellb
output.header.mapc = original.header.mapc
output.header.mapr = original.header.mapr
output.header.maps = original.header.maps
output.header.ispg = original.header.ispg
output.header.nsymbt = original.header.nsymbt
output.set_extended_header(original.extended_header)
output.header.exttyp = original.header.exttyp
output.header.nversion = original.header.nversion
output.header.origin = original.header.origin
output.header.map = original.header.map
output.header.machst = original.header.machst
output.header.rms = original.header.rms
output.header.nlabl = original.header.nlabl
output.header.label = original.header.label
output.close()
def design_matrix_classification(N, kernel_mode, X):
ret = np.ndarray(shape=(N, N))
for i in range(0, N):
for j in range(0, N):
if (kernel_mode == "polynomial"):
ret[i, j] = kernel.polynomial_kernel(X[i], X[j - 1])
elif (kernel_mode == "linear_spline"):
ret[i, j] = kernel.linear_spline(X[i], X[j - 1])
elif (kernel_mode == "gaussian"):
ret[i, j] = kernel.gaussian(X[i], X[j - 1])
elif (kernel_mode == "rbf"):
ret[i, j] = kernel.rbf(X[i], X[j - 1])
return ret
def Relevance_Vector_Classification_Prediction (Xtest, relevance_vectors, weightMaxPosteriori, kernel_choice):
Psum = 0
res = []
for xi in range(len(Xtest)):
for ri in range (len(relevance_vectors)):
if kernel_choice=="gaussian":
Psum+=weightMaxPosteriori[ri+1]*(kernel.gaussian(Xtest[xi], relevance_vectors[ri])) + weightMaxPosteriori[0]
elif kernel_choice=="linear":
Psum+=weightMaxPosteriori[ri+1]*(kernel.linear_kernel(Xtest[xi], relevance_vectors[ri])) + weightMaxPosteriori[0]
elif kernel_choice=="polynomial":
Psum+=weightMaxPosteriori[ri+1]*(kernel.polynomial_kernel(Xtest[xi], relevance_vectors[ri])) + weightMaxPosteriori[0]
elif kernel_choice=="linear_spline":
Psum+=weightMaxPosteriori[ri+1]*(kernel.linear_spline(Xtest[xi], relevance_vectors[ri])) + weightMaxPosteriori[0]
y = sigmoid_function(Psum)
if y >0.5:
res.append(1)
elif y<= 0.5:
res.append(0)
return res
X = np.append(np.ones((1, I_0 + I_1)), X, axis=0)
w = np.append(np.zeros((I_0, 1)), np.ones((I_1, 1)), axis=0)
var_prior = 6
X_test = np.arange(-5, 5, 0.1)
X_test = np.append(
np.ones((1, X_test.size)),
X_test.reshape(1, X_test.size),
axis=0
)
p_lambdas = [0.3, 1, 5, 15]
for index, p_lambda in enumerate(p_lambdas):
initial_psi = np.zeros((X.shape[1], 1))
predictions, psi = classification.fit_gaussian_process(
X, w, var_prior, X_test, initial_psi,
lambda x_i, x_j: kernel.gaussian(x_i, x_j, p_lambda)
)
plt.subplot(2, 2, index + 1)
plt.plot(np.arange(-5, 5, 0.1), predictions)
plt.scatter(class_0, np.zeros((1, I_0)), 50, c="r", edgecolors="k")
plt.scatter(class_1, np.zeros((1, I_1)), 50, c="g", edgecolors="k")
plt.axis([-5, 5, -0.1, 1.1])
plt.figure("2D Gaussian process classification")
granularity = 100
a = -5
b = 5
domain = np.linspace(a, b, granularity)
X, Y = np.meshgrid(domain, domain)
x = X.reshape((1, X.size))
# Prepare the training input
X_train = np.append(np.ones((1, X_data.shape[0])),
X_data[:, 0].reshape((1, X_data.shape[0])),
axis=0)
w = X_data[:, 1].reshape((X_data.shape[0], 1))
var_prior = 6
X_test = domain.reshape((1, granularity))
X_test = np.append(np.ones((1, granularity)), X_test, axis=0)
# Train 6 Gaussian process regression model for different values for nu
plt.figure("Fit Gaussian process regression")
nus = [0.5, 0.7, 0.9, 1.5, 2, 3]
for nu_index, nu in enumerate(nus):
mu_test, var_test = regression.fit_gaussian_process(
X_train, w, var_prior, X_test,
lambda x_i, x_j: kernel.gaussian(x_i, x_j, nu))
Z = np.zeros((granularity, granularity))
for j in range(granularity):
mu = mu_test[j, 0]
var = var_test[j, 0]
for i in range(granularity):
ww = domain[i]
Z[i, j] = gaussian(ww, mu, var)
plt.subplot(3, 2, nu_index + 1)
plt.pcolor(X, Y, Z)
plt.scatter(X_data[:, 0], X_data[:, 1], edgecolors="w")
plt.axis([-5, 5, -5, 5])
plt.show()
'''Use case: MNIST'''
import mnist
import torch
import kernel
import eigenpro
n_class = 10
(x_train, y_train), (x_test, y_test) = mnist.load()
x_train, y_train, x_test, y_test = x_train.astype('float32'), \
y_train.astype('float32'), x_test.astype('float32'), y_test.astype('float32')
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kernel_fn = lambda x,y: kernel.gaussian(x, y, bandwidth=5)
model = eigenpro.FKR_EigenPro(kernel_fn, x_train, n_class, device=device)
_ = model.fit(x_train, y_train, x_test, y_test, epochs=[1, 2, 5], mem_gb=12)
