def train_model(N, sparsity, eta):
W = np.random.normal(0, 1, [N, N])
W = W * (RHO / np.max(np.abs(np.linalg.eig(W)[0])))
WI = random.uniform(-TAU, TAU, N)
# make sparse
sparsity = 0.99
for i in range(N):
for j in range(N):
if i != j and random.random() < sparsity:
W[i, j] = 0.0
W = W * (RHO / np.max(np.abs(np.linalg.eig(W)[0])))
print(1.0 - np.count_nonzero(W) / (N * N))
mc_max = -inf
W = learn_orthonormal(W, eta)
while True:
print(1.0 - np.count_nonzero(W) / (N * N))
W = learn_orthonormal(W, eta)
last_mc = mc(WI, W, iters_skip=N, iters_train=10*N, iters_test=1000)
if mc_max < last_mc:
mc_max = last_mc
if mc_max > last_mc:
return W
def train_model(N, eta):
W = np.random.normal(0, 1, [N, N])
W = W * (RHO / np.max(np.abs(np.linalg.eig(W)[0])))
WI = random.uniform(-TAU, TAU, N)
mc_max = -inf
opt_iters = 0
W = learn_orthonormal(W, eta)
while True:
# W = learn_orthogonal(W, eta)
W = learn_orthonormal(W, eta)
last_mc = mc(WI, W, iters_skip=N, iters_train=10*N, iters_test=1000)
print('N={}, eta={}, mc={}'.format(N, eta, last_mc))
if mc_max < last_mc:
mc_max = last_mc
if mc_max > last_mc:
print()
return mc_max, opt_iters
opt_iters += 1
def train_model(N, eta):
W = np.random.normal(0, 1, [N, N])
W = W * (RHO / np.max(np.abs(np.linalg.eig(W)[0])))
WI = random.uniform(-TAU, TAU, N)
mc_max = -inf
opt_iters = 0
W = learn_orthonormal(W, eta)
while True:
# W = learn_orthogonal(W, eta)
W = learn_orthonormal(W, eta)
last_mc = mc(WI, W, iters_skip=N, iters_train=10 * N, iters_test=1000)
print('N={}, eta={}, mc={}'.format(N, eta, last_mc))
if mc_max < last_mc:
mc_max = last_mc
if mc_max > last_mc:
print()
return mc_max, opt_iters
opt_iters += 1
def generate_data():
numdata = 30000
x = np.random.uniform(eta_min, eta_max, numdata)
y = np.random.uniform(xi_min, xi_max, numdata)
z = np.zeros(numdata)
for i in range(numdata):
eta = x[i]
decay = y[i]
WG = np.random.normal(0, sigma, [N, N])
WI = np.random.uniform(-tau, tau, N)
for it in range(ORTHO_ITERATIONS):
WG = learn_orthonormal(WG, eta)
eta = eta * decay
z[i], _ = measure_mc(WG, WI, False)
np.save('x', x)
np.save('y', y)
np.save('z', z)
print(z)
return x, y, z
def generate_data():
numdata = 30000
x = np.random.uniform(eta_min, eta_max, numdata)
y = np.random.uniform(xi_min, xi_max, numdata)
z = np.zeros(numdata)
for i in range(numdata):
eta = x[i]
decay = y[i]
WG = np.random.normal(0, sigma, [N, N])
WI = np.random.uniform(-tau, tau, N)
for it in range(ORTHO_ITERATIONS):
WG = learn_orthonormal(WG, eta)
eta = eta * decay
z[i], _ = measure_mc(WG, WI, False)
np.save('x', x)
np.save('y', y)
np.save('z', z)
print(z)
return x, y, z
for rsi, N in enumerate(reservoir_sizes):
print(N)
mc_before = np.zeros(INSTANCES)
mc_after = np.zeros(INSTANCES)
le_before = np.zeros(INSTANCES)
le_after = np.zeros(INSTANCES)
for inst in range(INSTANCES):
W = random.normal(0, 1, [N, N])
W = W * (rho / np.max(np.abs(np.linalg.eig(W)[0])))
WI = random.uniform(-tau, tau, N)
mc_before[inst], le_before[inst] = measure_mc(W, WI)
eta = eta_0
for _ in range(ORTHOPROCESS_ITERATIONS):
W = learn_orthonormal(W, eta)
eta = eta * 0.9
mc_after[inst], le_after[inst] = measure_mc(W, WI)
mcb_mean[rsi] = np.average(mc_before)
mcb_std[rsi] = np.std(mc_before)
leb_mean[rsi] = np.average(le_before)
leb_std[rsi] = np.std(le_before)
mca_mean[rsi] = np.average(mc_after)
mca_std[rsi] = np.std(mc_after)
lea_mean[rsi] = np.average(le_after)
lea_std[rsi] = np.std(le_after)
np.save('mcbm', mcb_mean)
WG = W
mc_gon = np.zeros(ORTHO_ITERATIONS + 1)
les = []
mc_gon[0], le = measure_mc(WG, WI, True)
les.append(le)
eigen_values = np.zeros([ORTHO_ITERATIONS + 1, N])
singular_values = np.zeros([ORTHO_ITERATIONS + 1, N])
eigen_values[0, :] = np.sort(np.abs(np.linalg.eig(W)[0]))
singular_values[0, :] = np.linalg.svd(W, compute_uv=False)
eta = eta_0
for it in range(ORTHO_ITERATIONS):
WG = learn_orthonormal(WG, eta)
mc_gon[it + 1], le = measure_mc(WG, WI, True)
les.append(le)
eigen_values[it + 1, :] = np.sort(np.abs(np.linalg.eig(WG)[0]))
singular_values[it + 1, :] = np.linalg.svd(WG, compute_uv=False)
eta = eta * 0.95
MCs.append(mc_gon)
EVs.append(eigen_values)
SVs.append(singular_values)
LEs.append(les)
print()
np.save('mc', MCs)
def measure_og(W):
return orthogonality(W)
OTHS = np.zeros([LINES, ITERATIONS + 1])
LES = np.zeros([LINES, ITERATIONS + 1])
for l in range(LINES):
W = random.normal(0, sigma, [q, q])
WI = random.uniform(-tau, tau, q)
mc, le = measure_mc(W, WI)
og = measure_og(W)
OTHS[l, 0] = og
LES[l, 0] = le
eta = eta_0
for it in range(ITERATIONS):
W = learn_orthonormal(W, eta)
mc, le = measure_mc(W, WI)
og = measure_og(W)
OTHS[l, it + 1] = og
LES[l, it + 1] = le
eta = eta * 0.95
save('oths', OTHS)
save('les', LES)
WG = W
mc_gon = np.zeros(ORTHO_ITERATIONS + 1)
les = []
mc_gon[0], le = measure_mc(WG, WI, True)
les.append(le)
eigen_values = np.zeros([ORTHO_ITERATIONS + 1, N])
singular_values = np.zeros([ORTHO_ITERATIONS + 1, N])
eigen_values[0, :] = np.sort(np.abs(np.linalg.eig(W)[0]))
singular_values[0, :] = np.linalg.svd(W, compute_uv=False)
eta = eta_0
for it in range(ORTHO_ITERATIONS):
WG = learn_orthonormal(WG, eta)
mc_gon[it + 1], le = measure_mc(WG, WI, True)
les.append(le)
eigen_values[it + 1, :] = np.sort(np.abs(np.linalg.eig(WG)[0]))
singular_values[it + 1, :] = np.linalg.svd(WG, compute_uv=False)
eta = eta * 0.95
MCs.append(mc_gon)
EVs.append(eigen_values)
SVs.append(singular_values)
LEs.append(les)
print()
np.save('mc', MCs)
WN = W
mc_gon = np.zeros(ORTHO_ITERATIONS + 1)
les = []
mc_gon[0], le = measure_mc(WN, WI, True)
les.append(le)
eigen_values = np.zeros([ORTHO_ITERATIONS + 1, N])
singular_values = np.zeros([ORTHO_ITERATIONS + 1, N])
eigen_values[0, :] = np.sort(np.abs(np.linalg.eig(W)[0]))
singular_values[0, :] = np.linalg.svd(W, compute_uv=False)
eta = eta_0
for it in range(ORTHO_ITERATIONS):
WN = learn_orthonormal(WN, eta)
calc_lyapunov = (it + 1) % 10 == 0
mc_gon[it + 1], le = measure_mc(WN, WI, calc_lyapunov)
les.append(le)
eigen_values[it + 1, :] = np.sort(np.abs(np.linalg.eig(WN)[0]))
singular_values[it + 1, :] = np.linalg.svd(WN, compute_uv=False)
eta = eta * 0.95
MCs.append(mc_gon)
EVs.append(eigen_values)
SVs.append(singular_values)
LEs.append(les)
print()
eigenvaluesG = np.zeros([ORTHO_ITERATIONS + 1, N])
singular_valuesG = np.zeros([ORTHO_ITERATIONS + 1, N])
eigenvaluesN = np.zeros([ORTHO_ITERATIONS + 1, N])
singular_valuesN = np.zeros([ORTHO_ITERATIONS + 1, N])
mc_gon[0] = measure_mc(WG, WI)
mc_nor[0] = measure_mc(WN, WI)
eigenvaluesG[0, :] = np.sort(np.abs(np.linalg.eig(W)[0]))
singular_valuesG[0, :] = np.linalg.svd(W, compute_uv=False)
eigenvaluesN[0, :] = np.sort(np.abs(np.linalg.eig(W)[0]))
singular_valuesN[0, :] = np.linalg.svd(W, compute_uv=False)
for it in range(ORTHO_ITERATIONS):
print('\riteration', it, 'of', ORTHO_ITERATIONS, end='')
WG = learn_orthogonal(WG, eta)
WN = learn_orthonormal(WN, eta)
mc_gon[it + 1] = measure_mc(WG, WI)
mc_nor[it + 1] = measure_mc(WN, WI)
eigenvaluesG[it + 1, :] = np.sort(np.abs(np.linalg.eig(WG)[0]))
singular_valuesG[it + 1, :] = np.linalg.svd(WG, compute_uv=False)
eigenvaluesN[it + 1, :] = np.sort(np.abs(np.linalg.eig(WN)[0]))
singular_valuesN[it + 1, :] = np.linalg.svd(WN, compute_uv=False)
print()
def replot():
xs = range(ORTHO_ITERATIONS + 1)
