epochs = 20000
N = 10000
np.random.seed(123)
samples = uniform(n=N)
print 'Neural Gas'
np.random.seed(123)
ng = NG((n,n,2))
ng.learn(samples,epochs)
print 'Self-Organizing Map'
np.random.seed(123)
som = SOM((n,n,2))
som.learn(samples,epochs)
print 'Dynamic Self-Organizing Map'
np.random.seed(123)
dsom = DSOM((n,n,2), elasticity=1.75)
dsom.learn(samples,epochs)
fig = plt.figure(figsize=(21,8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(1,3,1)
ng.plot(axes)
axes = fig.add_subplot(1,3,2)
som.plot(axes)
axes = fig.add_subplot(1,3,3)
dsom.plot(axes)
fig.savefig('uniform.png',dpi=150)
samples = np.zeros((N, 2))
#samples[:,0] = np.array([0.1, 0.1, 0.5, 0.9])
#samples[:,1] = np.array([0.1, 0.9, 0.5, 0.9])
samples[:, 0] = np.array([0.1, 0.1, 0.5, 0.9, 0.9])
samples[:, 1] = np.array([0.1, 0.9, 0.5, 0.1, 0.9])
print 'Neural Gas'
np.random.seed(2)
ng = NG((n, n, 2))
ng.learn(samples, epochs)
print 'Self-Organizing Map'
np.random.seed(2)
som = SOM((n, n, 2))
som.learn(samples, epochs)
print 'Dynamic Self-Organizing Map'
np.random.seed(2)
dsom = DSOM((n, n, 2), elasticity=1.0, lrate=0.1)
#dsom = DSOM((n,n,2), elasticity=1.75)
dsom.learn(samples, epochs)
fig = plt.figure(figsize=(21, 8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(1, 3, 1)
ng.plot(axes)
axes = fig.add_subplot(1, 3, 2)
som.plot(axes)
axes = fig.add_subplot(1, 3, 3)
dsom.plot(axes)
fig.savefig('false.png', dpi=150)
np.random.seed(123)
samples = image(filename='lena.png', shape=(p,p), n=N)
print 'Neural Gas'
np.random.seed(123)
ng = NG((n,n,p*p), init_method='fixed')
ng.learn(samples,epochs)
print 'Self-Organizing Map'
np.random.seed(123)
som = SOM((n,n,p*p), init_method='fixed')
som.learn(samples,epochs)
print 'Dynamic Self-Organizing Map'
np.random.seed(123)
dsom = DSOM((n,n,p*p), elasticity=0.5, init_method='fixed')
dsom.learn(samples,epochs)
# fig = plt.figure(figsize=(10,10))
# axes = plt.subplot(111, frameon=False)
# plot(dsom,n,p)
fig = plt.figure(figsize=(21,8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(131, frameon=False)
plot(ng,n,p)
axes = plt.subplot(132, frameon=False)
plot(som,n,p)
axes = plt.subplot(133, frameon=False)
plot(dsom,n,p)
fig.savefig('image.png',dpi=150)
for i in range(N):
samples[i] = gaussian(shape=(p, p), sigma=(.5, 2),
theta=T[i]).flatten()
print 'Neural Gas'
np.random.seed(123)
ng = NG((n, n, p * p), init_method='fixed')
ng.learn(samples, epochs, noise=0.1)
print 'Self-Organizing Map'
np.random.seed(123)
som = SOM((n, n, p * p), init_method='fixed')
som.learn(samples, epochs, noise=0.1)
print 'Dynamic Self-Organizing Map'
np.random.seed(123)
dsom = DSOM((n, n, p * p), elasticity=1.5, init_method='fixed')
dsom.learn(samples, epochs, noise=0.1)
fig = plt.figure(figsize=(21, 8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(131, frameon=False)
plot(ng, n, p)
axes = plt.subplot(132, frameon=False)
plot(som, n, p)
axes = plt.subplot(133, frameon=False)
plot(dsom, n, p)
fig.savefig('gaussian-filters.png', dpi=150)
#plt.show()
#samples[:,1] = np.array([0.1, 0.9, 0.5, 0.9])
samples[:,0] = np.array([0.1, 0.1, 0.5, 0.9, 0.9])
samples[:,1] = np.array([0.1, 0.9, 0.5, 0.1, 0.9])
print 'Neural Gas'
np.random.seed(2)
ng = NG((n,n,2))
ng.learn(samples,epochs)
print 'Self-Organizing Map'
np.random.seed(2)
som = SOM((n,n,2))
som.learn(samples,epochs)
print 'Dynamic Self-Organizing Map'
np.random.seed(2)
dsom = DSOM((n,n,2), elasticity=1.0, lrate=0.1)
#dsom = DSOM((n,n,2), elasticity=1.75)
dsom.learn(samples,epochs)
fig = plt.figure(figsize=(21,8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(1,3,1)
ng.plot(axes)
axes = fig.add_subplot(1,3,2)
som.plot(axes)
axes = fig.add_subplot(1,3,3)
dsom.plot(axes)
fig.savefig('false.png',dpi=150)
v = x/float(mag)
b = np.array([v+.2, np.sin(7*(v+.2))/5 + np.random.random_sample()/3+ 0.3])
s3 = np.row_stack((s3,b))
samps = [s1,s2,s3]
n = len(samps)
epcs = [epochs//n for x in range(n)]
print 'Self-Organizing Map'
np.random.seed(12345)
som = SOM((size,size,2))
som.learn(samps,epcs)
print 'Dynamic Self-Organizing Map'
np.random.seed(12345)
dsom = DSOM((size,size,2), elasticity=2)
dsom.learn(samps,epcs)
fig = plt.figure(figsize=(16,7))
fig.patch.set_alpha(1.0)
axes = plt.subplot(1,2,1)
som.plot(axes)
axes = plt.subplot(1,2,2)
dsom.plot(axes)
fig.savefig('dynamicA.png',dpi=150)
fig.clf()
axes = plt.subplot(1,2,1)
som.plot_dist(axes)
rho = 0.5
u = np.random.uniform(low=0, high=1, size=N//2)
v = np.random.uniform(low=0, high=1, size=N//2)
theta = 2*np.pi*u
phi = np.arccos(2*v-1)
spheres[:N//2,0] = rho*np.cos(theta)*np.sin(phi)+.5
spheres[:N//2,1] = rho*np.sin(theta)*np.sin(phi)+.0
spheres[:N//2,2] = rho*np.cos(phi)+.5
spheres[N//2:,0] = rho*np.cos(theta)*np.sin(phi)+.5
spheres[N//2:,1] = rho*np.sin(theta)*np.sin(phi)+1.0
spheres[N//2:,2] = rho*np.cos(phi)+.5
samples = sphere
np.random.seed(123)
net = DSOM((n,n,3), elasticity=1.0, init_method='fixed')
I = np.random.randint(0,samples.shape[0], epochs)
bar = ProgressBar(widgets=[Percentage(), Bar()], maxval=epochs).start()
plotfile = '/tmp/plot.txt'
datafile = '/tmp/data.txt'
rot_x, rot_z = 65,225
for i in range(epochs):
if i == (epochs//2):
samples = spheres
I = np.random.randint(0,samples.shape[0], epochs)
if i%5 == 0:
rot_x = 20+(1+np.cos(i/float(epochs)*4*np.pi))*45
rot_z = (rot_z+1) % 360
print 'Neural gas'
np.random.seed(12345)
ng = NG((size,size,2))
ng.learn([samples_1, samples_2, samples_3, samples_4],
[2*epochs//8, 2*epochs//8, 2*epochs//8, 2*epochs//8])
print 'Self-Organizing Map'
np.random.seed(12345)
som = SOM((size,size,2))
som.learn([samples_1, samples_2, samples_3, samples_4],
[2*epochs//8, 2*epochs//8, 2*epochs//8, 2*epochs//8])
print 'Dynamic Self-Organizing Map'
np.random.seed(12345)
dsom = DSOM((size,size,2), elasticity=2.5)
dsom.learn([samples_1, samples_2, samples_3, samples_4],
[2*epochs//8, 2*epochs//8, 2*epochs//8, 2*epochs//8])
fig = plt.figure(figsize=(21,8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(1,3,1)
ng.plot(axes)
axes = plt.subplot(1,3,2)
som.plot(axes)
axes = plt.subplot(1,3,3)
dsom.plot(axes)
fig.savefig('dynamic.png',dpi=150)
#plt.show()
epochs = 1000
N = 2
#np.random.seed(123)
samples = uniform(n=N)
samples[:,0] = [0.0,1.0]
samples[:,1] = [0.5,0.5]
fig = plt.figure(figsize=(12,9))
p = 200
X,Y = np.zeros((p,)), np.zeros((p,))
for n in [2,4,6,8,10]:
for i in range(p):
elasticity = 0.5 +i*2.0/p
np.random.seed(123)
dsom = DSOM((n,1,2), elasticity=elasticity)
dsom.codebook[...] = [0.5,0.5]
dsom.learn(samples,epochs,show_progress=False)
x1,x2 = dsom.codebook[0,0][0], dsom.codebook[-1,0][0]
X[i],Y[i] = elasticity, 1-np.sqrt((x2-x1)**2)
print n, elasticity, Y[i]
plt.plot(X,Y,lw=2)
plt.legend(('n=2', 'n=4', 'n=6', 'n=8', 'n=10'), 'upper left')
plt.xlabel('Elasticity', fontsize=20)
plt.ylabel('Distortion', fontsize=20)
plt.show()
np.random.seed(123)
samples = np.zeros((N, 3))
samples[:, 0] = np.random.uniform(low=0, high=1, size=N)
samples[:, 1] = np.random.uniform(low=0, high=1, size=N)
samples[:, 2] = np.random.uniform(low=0, high=1, size=N)
for i in range(N):
index = int(np.random.random() * 3)
value = int(np.random.random() * 2)
samples[i, index] = value
x = samples[:, 0]
y = samples[:, 1]
z = samples[:, 2]
# Learn
np.random.seed(123)
dsom = DSOM((n, n, 3), elasticity=1.0, init_method='fixed')
dsom.learn(samples, epochs)
fig = plt.figure(figsize=(10, 10))
axes = Axes3D(fig)
axes.scatter(x, y, z)
C = dsom.codebook
Cx, Cy, Cz = C[..., 0], C[..., 1], C[..., 2]
for i in range(C.shape[0]):
axes.plot(Cx[i, :], Cy[i, :], Cz[i, :], 'k', alpha=0.85, lw=1.5)
for i in range(C.shape[1]):
axes.plot(Cx[:, i], Cy[:, i], Cz[:, i], 'k', alpha=0.85, lw=1.5)
axes.scatter(Cx.flatten(),
Cy.flatten(),
Cz.flatten(),
s=50,
for i in range(N):
samples[i] = gaussian(shape=(p,p), sigma=(.5,2), theta=T[i]).flatten()
print 'Neural Gas'
np.random.seed(123)
ng = NG((n,n,p*p), init_method='fixed')
ng.learn(samples,epochs, noise=0.1)
print 'Self-Organizing Map'
np.random.seed(123)
som = SOM((n,n,p*p), init_method='fixed')
som.learn(samples,epochs, noise=0.1)
print 'Dynamic Self-Organizing Map'
np.random.seed(123)
dsom = DSOM((n,n,p*p), elasticity=1.5, init_method='fixed')
dsom.learn(samples,epochs, noise=0.1)
fig = plt.figure(figsize=(21,8))
fig.patch.set_alpha(0.0)
axes = plt.subplot(131, frameon=False)
plot(ng,n,p)
axes = plt.subplot(132, frameon=False)
plot(som,n,p)
axes = plt.subplot(133, frameon=False)
plot(dsom,n,p)
fig.savefig('gaussian-filters.png',dpi=150)
#plt.show()
n = 8
epochs = 20000
N = 10000
np.random.seed(123)
area_1 = np.pi * 0.5**2 - np.pi * 0.25**2
area_2 = np.pi * 0.25**2
n1 = int(area_1 * 25000)
n2 = int(area_2 * 25000)
samples = np.zeros((n1 + n2, 2))
samples[:n1] = ring(n=n1, radius=(0.25, 0.50))
samples[n1:] = ring(n=n2, radius=(0.00, 0.25))
print 'Dynamic Self-Organizing Map 1'
np.random.seed(123)
dsom1 = DSOM((n, n, 2), elasticity=1.25, init_method='fixed')
dsom1.learn(samples, epochs)
n1 = int(area_1 * 40000)
n2 = int(area_2 * 10000)
samples = np.zeros((n1 + n2, 2))
samples[:n1] = ring(n=n1, radius=(0.25, 0.50))
samples[n1:] = ring(n=n2, radius=(0.00, 0.25))
print 'Dynamic Self-Organizing Map 2'
np.random.seed(123)
dsom2 = DSOM((n, n, 2), elasticity=1.25, init_method='fixed')
dsom2.learn(samples, epochs)
n1 = int(area_1 * 10000)
n2 = int(area_2 * 40000)
samples = np.zeros((n1 + n2, 2))
N = 2
# np.random.seed(123)
samples = uniform(n=N)
samples[:, 0] = [0.0, 1.0]
samples[:, 1] = [0.5, 0.5]
fig = plt.figure(figsize=(15, 6))
p = 120
X, Y = np.zeros((p,)), np.zeros((p,))
lrate = 0.1
for s in [0.5, 0.8, 0.85, 0.9, 0.95]:
for i in range(p):
elasticity = 1.0 + i * (3.0 / p)
np.random.seed(123)
dsom = DSOM((n, 1, 2), elasticity=elasticity, lrate=lrate)
# dsom.codebook[...] = [0.5,0.5]
dsom.codebook[0] = [s, 0.5]
dsom.codebook[1] = [1 - s, 0.5]
dsom.learn(samples, epochs, show_progress=False)
x1, x2 = dsom.codebook[0, 0][0], dsom.codebook[-1, 0][0]
X[i], Y[i] = elasticity, 1 - np.sqrt((x2 - x1) ** 2)
print s, elasticity, Y[i]
plt.plot(X, Y, lw=2)
plt.legend(
(
r"$x_0 = 1-y_0 = 0.50$",
r"$x_0 = 1-y_0 = 0.20$",
r"$x_0 = 1-y_0 = 0.15$",
r"$x_0 = 1-y_0 = 0.10$",
r"$x_0 = 1-y_0 = 0.05$",
n = 8
epochs = 20000
N = 10000
np.random.seed(123)
area_1 = np.pi*0.5**2 - np.pi*0.25**2
area_2 = np.pi*0.25**2
n1 = int(area_1*25000)
n2 = int(area_2*25000)
samples = np.zeros((n1+n2,2))
samples[:n1] = ring(n=n1, radius=(0.25,0.50))
samples[n1:] = ring(n=n2, radius=(0.00,0.25))
print 'Dynamic Self-Organizing Map 1'
np.random.seed(123)
dsom1 = DSOM((n,n,2), elasticity=1.25, init_method='fixed')
dsom1.learn(samples,epochs)
n1 = int(area_1*40000)
n2 = int(area_2*10000)
samples = np.zeros((n1+n2,2))
samples[:n1] = ring(n=n1, radius=(0.25,0.50))
samples[n1:] = ring(n=n2, radius=(0.00,0.25))
print 'Dynamic Self-Organizing Map 2'
np.random.seed(123)
dsom2 = DSOM((n,n,2), elasticity=1.25, init_method='fixed')
dsom2.learn(samples,epochs)
n1 = int(area_1*10000)
n2 = int(area_2*40000)
samples = np.zeros((n1+n2,2))
N = 2
#np.random.seed(123)
samples = uniform(n=N)
samples[:, 0] = [0.0, 1.0]
samples[:, 1] = [0.5, 0.5]
fig = plt.figure(figsize=(15, 6))
p = 120
X, Y = np.zeros((p, )), np.zeros((p, ))
lrate = 0.1
for s in [0.5, 0.8, 0.85, 0.9, 0.95]:
for i in range(p):
elasticity = 1.0 + i * (3.0 / p)
np.random.seed(123)
dsom = DSOM((n, 1, 2), elasticity=elasticity, lrate=lrate)
#dsom.codebook[...] = [0.5,0.5]
dsom.codebook[0] = [s, 0.5]
dsom.codebook[1] = [1 - s, 0.5]
dsom.learn(samples, epochs, show_progress=False)
x1, x2 = dsom.codebook[0, 0][0], dsom.codebook[-1, 0][0]
X[i], Y[i] = elasticity, 1 - np.sqrt((x2 - x1)**2)
print s, elasticity, Y[i]
plt.plot(X, Y, lw=2)
plt.legend((r'$x_0 = 1-y_0 = 0.50$', r'$x_0 = 1-y_0 = 0.20$',
r'$x_0 = 1-y_0 = 0.15$', r'$x_0 = 1-y_0 = 0.10$',
r'$x_0 = 1-y_0 = 0.05$'), 'upper right')
plt.xlabel('Elasticity', fontsize=16)
plt.ylabel('Error', fontsize=16)
plt.title('Error as a function of initial conditions and elasticity',
fontsize=20)
# Pick N points on a cube uniformly
np.random.seed(123)
samples = np.zeros((N, 3))
samples[:, 0] = np.random.uniform(low=0, high=1, size=N)
samples[:, 1] = np.random.uniform(low=0, high=1, size=N)
samples[:, 2] = np.random.uniform(low=0, high=1, size=N)
for i in range(N):
index = int(np.random.random() * 3)
value = int(np.random.random() * 2)
samples[i, index] = value
x = samples[:, 0]
y = samples[:, 1]
z = samples[:, 2]
np.random.seed(123)
net = DSOM((n, n, 3), elasticity=1.0, init_method='fixed')
I = np.random.randint(0, samples.shape[0], epochs)
bar = ProgressBar(widgets=[Percentage(), Bar()], maxval=epochs).start()
plotfile = '/tmp/plot'
datafile = '/tmp/data.txt'
rot_x, rot_z = 65, 225
for i in range(epochs):
if i % 5 == 0:
rot_x = 20 + (1 + np.cos(i / float(epochs) * 4 * np.pi)) * 45
rot_z = (rot_z + 1) % 360
filename = '/tmp/image-%05d' % i
file = open(plotfile, 'w')
file.write('''set parametric\n''')
file.write('''set hidden3d\n''')
file.write('''unset key\n''')
file.write('''unset border\n''')
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from network import NG, SOM, DSOM
from distribution import uniform, normal, ring
n = 8
epochs = 20000
N = 10000
np.random.seed(12345)
samples = normal(n=N)
print 'Dynamic Self-Organizing Map, elasticity = 1.0'
np.random.seed(12345)
dsom1 = DSOM((n, n, 2), elasticity=1.0, init_method='regular')
dsom1.learn(samples, epochs)
print 'Dynamic Self-Organizing Map, elasticity = 2.0'
np.random.seed(12345)
dsom2 = DSOM((n, n, 2), elasticity=2.0, init_method='regular')
dsom2.learn(samples, epochs)
print 'Dynamic Self-Organizing Map, elasticity = 3.0'
np.random.seed(12345)
dsom3 = DSOM((n, n, 2), elasticity=3.0, init_method='regular')
dsom3.learn(samples, epochs)
fig = plt.figure(figsize=(21, 8))
fig.patch.set_alpha(0.0)
