def TrackVelocity(x, k, vmax, acmax, Ta):
''' compute the velocity at each point along the track (given
already-computed curvatures) assuming a certain accelration profile '''
v = np.minimum(np.abs(acmax / k)**0.5, vmax)
# also compute arc distance between successive points in x given curvature
# k; for now we'll just use the linear distance though as it's close enough
s = np.abs(np.concatenate([x[1:] - x[:-1], x[:1] - x[-1:]]))
va = 0
T = 0
vout = []
# first pass is just to get the initial velocity
# let's assume it's zero
# for i in range(1, len(k)):
# va = va + (v[i] - va) / Ta
for i in range(0, len(k)):
a = (v[i] - va) / Ta # acceleration
dt = s[i] / (va + a/2) # time to reach next waypoint
va = np.minimum(va + dt * (v[i] - va) / Ta, v[i])
T += dt
vout.append(va)
return np.array(vout), T
def normalizing_flows(z_0, norm_flow_params):
'''
z_0: [n_samples, D]
u: [D,1]
w: [D,1]
b: [1]
'''
current_z = z_0
all_zs = []
all_zs.append(z_0)
for params_k in norm_flow_params:
z_0_mean = params_k[0]
a = np.abs(params_k[1])
b = params_k[2]
# m_x = -1. + np.log(1.+np.exp(np.dot(w.T,u)))
# u_k = u + (m_x - np.dot(w.T,u)) * (w/np.linalg.norm(w))
# print (a.shape)
# print (current_z.shape)
z_0_mean = np.reshape(z_0_mean, [len(current_z[0])])
h = 1./(a + np.abs(current_z - z_0_mean))
term1 = b * h * (current_z - z_0_mean)
current_z = current_z + term1
all_zs.append(current_z)
return current_z, all_zs
def intersect(sa, sb, image):
xlima, ylima = sa.bounding_boxes[image]
xlimb, ylimb = sb.bounding_boxes[image]
widtha, heighta = xlima[1] - xlima[0], ylima[1] - ylima[0]
widthb, heightb = xlimb[1] - xlimb[0], ylimb[1] - ylimb[0]
return (np.abs(xlima[0] - xlimb[0])*2 < (widtha + widthb)) and \
(np.abs(ylima[0] - ylimb[0])*2 < (heighta + heightb))
def PhotometricError(iref, inew, R, T, points, D):
# points is a tuple ([y], [x]); convert to homogeneous
siz = iref.shape
npoints = len(points[0])
f = siz[1] # focal length, FIXME
Xref = np.vstack(((points[1] - siz[1]*0.5) / f, # x
(siz[0]*0.5 - points[0]) / f, # y (left->right hand)
np.ones(npoints))) # z = 1
# this is confusingly written -- i am broadcasting the translation T to
# every column, but numpy broadcasting only works if it's rows, hence all
# the transposes
# print D * Xref
Xnew = (np.dot(so3.exp(R), (D * Xref)).T + T).T
# print Xnew
# right -> left hand projection
proj = Xnew[0:2] / Xnew[2]
p = (-proj[1]*f + siz[0]*0.5, proj[0]*f + siz[1]*0.5)
margin = 10 # int(siz[0] / 5)
inwindow_mask = ((p[0] >= margin) & (p[0] < siz[0]-margin-1) &
(p[1] >= margin) & (p[1] < siz[1]-margin-1))
npts_inw = sum(inwindow_mask)
if npts_inw < 10:
return 1e6, np.zeros(6 + npoints)
# todo: filter points which are now out of the window
oldpointidxs = (points[0][inwindow_mask],
points[1][inwindow_mask])
newpointidxs = (p[0][inwindow_mask], p[1][inwindow_mask])
origpointidxs = np.nonzero(inwindow_mask)[0]
E = InterpolatedValues(inew, newpointidxs) - iref[oldpointidxs]
# dE/dk ->
# d/dk r_p^2 = d/dk (Inew(w(r, T, D, p)) - Iref(p))^2
# = -2r_p dInew/dp dp/dw dw/dX dX/dk
# = -2r_p * g(w(r, T, D, p)) * dw(r, T, D, p)
# intensity gradients for each point
Ig = InterpolatedGradients(inew, newpointidxs)
# TODO: use tensors for this
# gradients for R, T, and D
gradient = np.zeros(6 + npoints)
for i in range(npts_inw):
# print 'newidx (y,x) = ', newpointidxs[0][i], newpointidxs[1][i]
# Jacobian of w
oi = origpointidxs[i]
Jw = dw(Xref[0][oi], Xref[1][oi], D[oi], R, T)
# scale back up into pixel space, right->left hand coords to get
# Jacobian of p
Jp = f * np.vstack((-Jw[1], Jw[0]))
# print origpointidxs[i], 'Xref', Xref[:, i], 'Ig', Ig[:, i], \
# 'dwdRz', Jw[:, 2], 'dpdRz', Jp[:, 2]
# full Jacobian = 2*E + Ig * Jp
J = np.sign(E[i]) * np.dot(Ig[:, i], Jp)
# print '2 E[i]', 2*E[i], 'Ig*Jp', np.dot(Ig[:, i], Jp)
gradient[:6] += J[:6]
# print J[:6]
gradient[6+origpointidxs[i]] += J[6]
print R, T, np.sum(np.abs(E)), npts_inw
# return ((0.2*(npoints - npts_inw) + np.dot(E, E)), gradient)
return np.sum(np.abs(E)) / (npts_inw), gradient / (npts_inw)
def compare_deltas(baseline=None, candidate=None, abs_tol=1e-5, rel_tol=0.01):
# TODO: maybe add relative tolerance check
epsilon = 1e-25
if baseline.shape != candidate.shape:
return False
diff_tensor = np.abs(baseline - candidate)
rel_tensor1 = diff_tensor / (np.abs(baseline) + 1e-25)
rel_tensor2 = diff_tensor / (np.abs(candidate) + 1e-25)
max_error = np.max(diff_tensor)
max_rel = max(np.max(rel_tensor1), np.max(rel_tensor2))
if max_error > abs_tol and max_rel > rel_tol:
return False
else:
return True
def _add_penalty(self, loss, w):
"""Apply regularization to the loss."""
if self.penalty == "l1":
loss += self.C * np.abs(w[:-1]).sum()
elif self.penalty == "l2":
loss += (0.5 * self.C) * (w[:-1] ** 2).mean()
return loss
def plot_projection(P, feature_names):
d = {}
import pandas as pd
for (i,v) in enumerate(P.T):
idxs = np.argsort(-np.abs(v))
print 'v_%d' % i
print pd.Series(v, index=feature_names).iloc[idxs].iloc[0:10]
def loss_func(self, w): loss = 0.5 * np.mean((self.train_y_ - np.dot(self.train_x_, w))**2) if self.penalty_ == "l1": # Lasso loss += self.alpha_ * np.sum(np.abs(w[:-1])) elif self.penalty_ == "l2": # Ridge loss += 0.5 * self.alpha_ * np.mean(w[:-1]**2) return loss
def SVGPathToTrackPoints(fname, spacing=TRACK_SPACING):
path = svgpathtools.parse_path(open(fname).read())
# input: svg based on a 100px / m image
def next_t(path, t, dist):
p = path.point(t)
L = path.length()
# t += 1.0 / np.abs(path.derivative(t))
itr = 0
while itr < 20:
itr += 1
p1 = path.point(t)
err = np.abs(p1 - p) - dist
d1 = path.derivative(t)
if np.abs(err) < 1e-5:
return t, p1, d1 / np.abs(d1)
derr = np.abs(d1) * L
# do a step in Newton's method (clipped because some of the
# gradients in the curve are really small)
t -= np.clip(err / derr, -1e-2, 1e-2)
t = np.clip(t, 0, 1)
return t, p, d1 / np.abs(d1)
d0 = path.derivative(0)
pts = [[path.point(0), d0 / np.abs(d0)]]
t = 0
while t < 1:
t, p, d = next_t(path, t, spacing)
pts.append([p, d])
return pts
def GetTimeSeries():
TS = []
ERR = []
GYRO = []
ACCEL = []
lastim = None
t0 = None
for msg in parse.ParseLog(open('../rustlerlog-BMPauR')):
if msg[0] == 'img':
_, ts, im = msg
if t0 is None:
t0 = ts
im = GammaCorrect(im)
if lastim is not None:
TS.append(ts - t0)
ERR.append(np.sum(np.abs(lastim-im)) / (320*240))
GYRO.append(GYRO[-1])
ACCEL.append(ACCEL[-1])
lastim = im
elif msg[0] == 'imu':
_, ts, gyro, mag, accel = msg
if t0 is None:
t0 = ts
TS.append(ts - t0)
GYRO.append(gyro)
ACCEL.append(accel)
if len(ERR):
ERR.append(ERR[-1])
else:
ERR.append(0)
return np.array(TS), np.array(ERR), np.array(GYRO), np.array(ACCEL)
def soft_thr(x, lambdaPar, lower=None, upper=None):
out = np.sign(x) * np.fmax(np.abs(x) - lambdaPar, 0)
if (lower != None):
out[out < lower] = 0.0
if (upper != None):
out[out > upper] = 0.0
return out
def TrackNormal(x):
xx = np.concatenate([x[-1:], x, x[:1]])
p0 = xx[:-2]
p2 = xx[2:]
T = p2 - p0 # track derivative
uT = np.abs(T)
return T / uT
def initialize(deep_map, X,num_pseudo_params):
smart_map = {}
for layer,layer_map in deep_map.iteritems():
smart_map[layer] = {}
for unit,gp_map in layer_map.iteritems():
smart_map[layer][unit] = {}
cov_params = gp_map['cov_params']
lengthscales = cov_params[1:]
if layer == 0:
pairs = itertools.combinations(X, 2)
dists = np.array([np.abs(p1-p2) for p1,p2 in pairs])
smart_lengthscales = np.array([np.log(np.median(dists[:,i])) for i in xrange(len(lengthscales))])
kmeans = KMeans(n_clusters = num_pseudo_params, init = 'k-means++')
fit = kmeans.fit(X)
smart_x0 = fit.cluster_centers_
#inds = npr.choice(len(X), num_pseudo_params, replace = False)
#smart_x0 = np.array(X)[inds,:]
smart_y0 = np.ndarray.flatten(smart_x0)
#smart_y0 = np.array(y)[inds]
smart_noise_scale = np.log(np.var(smart_y0))
else:
smart_x0 = gp_map['x0']
smart_y0 = np.ndarray.flatten(smart_x0[:,0])
smart_lengthscales = np.array([np.log(1) for i in xrange(len(lengthscales))])
smart_noise_scale = np.log(np.var(smart_y0))
gp_map['cov_params'] = np.append(cov_params[0],smart_lengthscales)
gp_map['x0'] = smart_x0
gp_map['y0'] = smart_y0
#gp_map['noise_scale'] = smart_noise_scale
smart_map[layer][unit] = gp_map
smart_params = pack_deep_params(smart_map)
return smart_params
def test_abs():
fun = lambda x: 3.0 * np.abs(x)
d_fun = grad(fun)
check_grads(fun, 1.1)
check_grads(fun, -1.1)
check_grads(d_fun, 1.1)
check_grads(d_fun, -1.1)
def taylor_sine(x):
ans = currterm = x
i = 0
while np.abs(currterm) > 0.001:
currterm = -currterm * x ** 2 / ((2 * i + 3) * (2 * i + 2))
ans = ans + currterm
i += 1
return ans
def fun(x):
curr = x
ans = curr
for i in xrange(1000):
curr = - curr * x**2 / ((2*i+3)*(2*i+2))
ans = ans + curr
if np.abs(curr) < 0.2: break
return ans
def next_t(path, t, dist):
p = path.point(t)
L = path.length()
# t += 1.0 / np.abs(path.derivative(t))
itr = 0
while itr < 20:
itr += 1
p1 = path.point(t)
err = np.abs(p1 - p) - dist
d1 = path.derivative(t)
if np.abs(err) < 1e-5:
return t, p1, d1 / np.abs(d1)
derr = np.abs(d1) * L
# do a step in Newton's method (clipped because some of the
# gradients in the curve are really small)
t -= np.clip(err / derr, -1e-2, 1e-2)
t = np.clip(t, 0, 1)
return t, p, d1 / np.abs(d1)
def hard_thr(x, lambdaPar, lower=None, upper=None):
out = np.copy(x)
out[np.abs(x) < lambdaPar] = 0.0
if (lower != None):
out[out < lower] = 0.0
if (upper != None):
out[out > upper] = 0.0
return out
def callback(weights, iter):
if iter % 10 == 0:
print "max of weights", np.max(np.abs(weights))
train_preds = undo_norm(pred_fun(weights, train_smiles))
cur_loss = loss_fun(weights, train_smiles, train_targets)
training_curve.append(cur_loss)
print "Iteration", iter, "loss", cur_loss, "train RMSE", rmse(train_preds, train_raw_targets),
if validation_smiles is not None:
validation_preds = undo_norm(pred_fun(weights, validation_smiles))
print "Validation RMSE", iter, ":", rmse(validation_preds, validation_raw_targets),
def test_pow():
fun = lambda x, y : to_scalar(x ** y)
d_fun_0 = lambda x, y : to_scalar(grad(fun, 0)(x, y))
d_fun_1 = lambda x, y : to_scalar(grad(fun, 1)(x, y))
make_positive = lambda x : np.abs(x) + 1.1 # Numeric derivatives fail near zero
for arg1, arg2 in arg_pairs():
arg1 = make_positive(arg1)
check_grads(fun, arg1, arg2)
check_grads(d_fun_0, arg1, arg2)
check_grads(d_fun_1, arg1, arg2)
def fun(x):
currterm = x
ans = currterm
for i in range(1000):
print(i, end=' ')
currterm = - currterm * x ** 2 / ((2 * i + 3) * (2 * i + 2))
ans = ans + currterm
if np.abs(currterm) < 0.2: break # (Very generous tolerance!)
return ans
def magcal_residual(MAG, a, mb):
""" residual from all observations given magnetometer eccentricity, bias,
gyro bias, and gyro scale"""
A = np.array([
[a[0], a[1], a[2]],
[0, a[3], a[4]],
[0, 0, a[5]]
])
mag = np.dot(MAG - mb, A)
return np.mean(np.abs(1 - np.einsum('ji,ji->j', mag, mag)))
def _apply_nonlinearity(self, nonlin, res):
if nonlin == 'none' or nonlin == 'linear':
return res
if nonlin == 'relu':
return 0.5 * (res + np.abs(res))
if nonlin == 'tanh':
return np.tanh(res)
if nonlin == 'softmax':
res -= res.max()
res = np.exp(res)
return res / res.sum()
raise Exception('unknown nonlinearity: "%s"' % nonlin)
def TrackCurvature(x):
# use quadratic b-splines at each point to estimate curvature
# i get almost the same formula i had before but it's off by a factor of 4!
xx = np.concatenate([x[-1:], x, x[:1]])
p0 = xx[:-2]
p1 = xx[1:-1]
p2 = xx[2:]
T = p2 - p0 # track derivative
uT = np.abs(T)
TT = 4*(p0 - 2*p1 + p2)
k = (np.real(T)*np.imag(TT) - np.imag(T)*np.real(TT)) / (uT**3)
return k
def callback(weights, iter):
if iter % 10 == 0:
print "max of weights", np.max(np.abs(weights))
# import pdb; pdb.set_trace()
train_preds = undo_norm(pred_fun(weights, train_smiles[:num_print_examples]))
cur_loss = loss_fun(weights, train_smiles[:num_print_examples], train_targets[:num_print_examples]) # V: refers to line number #78 i.e.
# def loss_fun(weights, smiles, targets) of build_vanilla_net.py
training_curve.append(cur_loss)
print "Iteration", iter, "loss", cur_loss,\
"train RMSE", rmse(train_preds, train_raw_targets[:num_print_examples]),
if validation_smiles is not None:
validation_preds = undo_norm(pred_fun(weights, validation_smiles))
print "Validation RMSE", iter, ":", rmse(validation_preds, validation_raw_targets),
def gradient_descent(objFunc, w): dfunc = grad(objFunc) lrate = 0.000000001 a = [] for i in range(8000): a.append(objFunc(w)) improv = lrate * dfunc(w) w = w - improv if i % 500 == 0: lrate = lrate/10.0 if len(a) > 3 and (np.abs(a[-1] - a[-2]) < 0.00001): break return w,a
def neural_net_train(features,labels,num_iter = 2000,opt_method = 'forward_backward') :
layer_sizes = [2,5,5,1]
l2_reg = 2.0
param_scale = 0.5
init_params = neural_net_init(param_scale,layer_sizes)
'''def plain_objective(params) :
return lms_loss(params,features,labels,l2_reg)'''
plain_objective = gen_objective(features,labels,l2_reg)
objective_grad = auto_grad(plain_objective)
print(" Iteration| Train accuracy")
optimized_params = init_params
gd_step = 0.2
for i in range(num_iter) :
if opt_method == 'forward_backward' :
optimized_params_ori = optimized_params
value_old = plain_objective(optimized_params)
flattened_grad,unflatten,x = flatten_func(objective_grad,optimized_params)
x -= flattened_grad(x) * gd_step
optimized_params = unflatten(x)
value_new = plain_objective(optimized_params)
if value_new < value_old :
gd_step *= 1.618
else :
gd_step *= 0.618
optimized_params = optimized_params_ori
elif opt_method == 'stepest' :
value_old = plain_objective(optimized_params)
flattened_grad,unflatten,x = flatten_func(objective_grad,optimized_params)
local_gd_step = gd_step
best_gd_step = 0.0
for j in range(10) :
x_test = x - flattened_grad(x) * local_gd_step
last_optimized_params = unflatten(x_test)
value_new = plain_objective(last_optimized_params)
if value_new < value_old :
best_gd_step = local_gd_step
local_gd_step *= 1.618
else :
local_gd_step *= 0.618
if auto_np.abs(best_gd_step - 0.0) < 0.00000000001 :
gd_step *= 0.618
x -= flattened_grad(x) * best_gd_step
optimized_params = unflatten(x)
print_perf(optimized_params,i,features,labels)
return optimized_params
def variational_log_density(params, samples):
'''
samples: [n_samples, D]
u: [D,1]
w: [D,1]
b: [1]
Returns: [num_samples]
'''
n_samples = len(samples)
d = len(samples[0])
mean = params[0]
log_std = params[1]
norm_flow_params = params[2]
# print (samples.shape)
# samples = sample_diag_gaussian(mean, log_std, num_samples, rs)
z_k, all_zs = normalizing_flows(samples, norm_flow_params)
logp_zk = logprob(z_k)
logp_zk = np.reshape(logp_zk, [n_samples, 1])
logq_z0 = diag_gaussian_log_density(samples, mean, log_std)
logq_z0 = np.reshape(logq_z0, [n_samples, 1])
sum_nf = np.zeros([n_samples,1])
for params_k in range(len(norm_flow_params)):
z_0_mean = norm_flow_params[params_k][0]
a = np.abs(norm_flow_params[params_k][1])
b = norm_flow_params[params_k][2]
# m_x = -1. + np.log(1.+np.exp(np.dot(w.T,u)))
# u_k = u + (m_x - np.dot(w.T,u)) * (w/np.linalg.norm(w))
# [n_samples, D]
# phi = np.dot((1.-np.tanh(np.dot(all_zs[params_k],w)+b)**2), w.T)
# [n_samples, 1]
current_z = all_zs[params_k]
z_0_mean = np.reshape(z_0_mean, [len(current_z[0])])
h = 1./(a + np.abs(current_z - z_0_mean))
h_prime = -1*(a+np.abs(current_z-z_0_mean))**2 * (np.abs(current_z)/current_z)
term1 = (1+b*h)**(d-1)
term2 = 1+ b * h + b * h_prime * np.abs(current_z-z_0_mean)
term3 = term1 * term2
sum_nf = np.log(np.abs(term3))
sum_nf += sum_nf
# return logq_z0 - sum_nf
print (logq_z0.shape)
log_qz = np.reshape(logq_z0 - sum_nf, [n_samples])
return log_qz
def QuadFitCurvatureMap(x):
curv = []
for i in range(len(x)):
# do a look-ahead quadratic fit, just like the car would do
pts = x[(np.arange(6) + i) % len(x)] / 100 # convert to meters
basis = (pts[1] - pts[0]) / np.abs(pts[1] - pts[0])
# project onto forward direction
pts = (np.conj(basis) * (pts - pts[0]))
p = np.polyfit(np.real(pts), np.imag(pts), 2)
curv.append(p[0] / 2)
return np.float32(curv)
def loss_func(self, w): prob = self.prob_func(w, self.train_x_) loss = np.array([0.0] * self.train_x_.shape[0]) for c in xrange(len(self.classes_)): loss += np.equal(self.train_y_, c) * prob[:,c] score = -np.sum(np.log(loss)) if self.penalty_ == "l1": # Lasso score += self.alpha_ * np.sum(np.abs(w)) elif self.penalty_ == "l2": # Ridge # ===FIXME=== # I don't know why scaler of L2 norm is "3.0" # according to definition of L2 norm, scaler must be "0.5" # but calc the same value with scikit-learn, scaler is "3.0" # (may be to avoid over learning?) score += 3.0 * self.alpha_ * np.mean(w**2) return score
def eigdecomp(self, x, k):
# self.hvp.update_x(x)
# d, U = doublePass(self.hvp, self.Omega, k, s=1, check=False)
J = self.J(x)
if len(np.shape(J)) == 1:
J = np.array([J])
Gauss_Newton = np.dot(np.dot(J.T, self.Gamma_noise_inv), J)
d, U = np.linalg.eigh(Gauss_Newton)
sort_perm = np.abs(d).argsort()
sort_perm = sort_perm[::-1]
d = d[sort_perm[:k]]
U = U[:, sort_perm[:k]]
return d, U
def _get_dx_wrt(dx, var, dx_scaling, dx_func=None):
"""Scale `dx` for a particular variable `var`."""
assert dx_scaling in [
"none",
"median",
"custom",
], "`dx_scaling` must be 'none' or 'median'."
if dx_scaling == "none":
dx_wrt = dx
elif dx_scaling == "median":
median_var = np.nanmedian(var)
if median_var == 0:
dx_wrt = dx
else:
dx_wrt = dx * np.abs(median_var)
elif dx_scaling == "custom":
dx_wrt = dx_func(var)
return dx_wrt
def one_of_K_code(arr):
"""
Make a one-of-K coding out of the numpy array.
For example, if arr = ([0, 1, 0, 2]), then return a 2d array of the form
[[1, 0, 0],
[0, 1, 0],
[1, 0, 0],
[0, 0, 1]]
"""
U = np.unique(arr)
n = len(arr)
nu = len(U)
X = np.zeros((n, nu))
for i, u in enumerate(U):
Ii = np.where(np.abs(arr - u) < 1e-8)
#ni = len(Ii)
X[Ii[0], i] = 1
return X
def nsp_ik(goal, q, nesterov=True):
def cost(q):
objective = (goal - fk(q[0:3], l1))
constraint = np.pi - q[2]
return objective.T @ objective + \
constraint.T * constraint
def g(q):
return fk(q[0:2], l1) - fk(q[3:], l2)
grad_cost = grad(cost)
constraint_jac = jacobian(g)
alpha = 0.01
nu = 0.9
max_itr = 100
v, n_itr = 0, 0
while np.any(np.abs(cost(q)) > EPS):
# get constraint jacobian
G = constraint_jac(q)
Ginv = np.linalg.pinv(G)
I = np.identity(Ginv.shape[0])
# matrix to projects into the null space of the constraint jacobian
nsp = (I - G.T @ Ginv.T).T
if nesterov:
v = nu * v + alpha * grad_cost(q - nu * v)
q -= nsp @ v
else:
q -= nsp @ (alpha * grad_cost(q))
draw_arm(q[0:3], l1, c1, swap=False)
draw_arm(q[3:], l2, c2, clear=False)
n_itr += 1
if n_itr > max_itr:
break
def test_archaic_and_pairwisediffs():
#logging.basicConfig(level=logging.DEBUG)
theta = 1
N_e = 1.0
join_time = 1.0
num_runs = 1000
def logit(p):
return np.log(p / (1. - p))
def expit(x):
return 1. / (1. + np.exp(-x))
n_bases = 1000
model = momi.DemographicModel(
N_e, muts_per_gen=theta/4./N_e/n_bases)
model.add_time_param(
"sample_t", random.uniform(0.001, join_time-.001) / join_time,
upper=join_time)
model.add_size_param("N", 1.0)
model.add_leaf("a", N="N")
model.add_leaf("b", t="sample_t", N="N")
model.move_lineages("a", "b", join_time)
data = model.simulate_data(length=n_bases,
recoms_per_gen=0,
num_replicates=num_runs,
sampled_n_dict={"a": 2, "b": 2})
model.set_data(data.extract_sfs(1),
use_pairwise_diffs=False,
mem_chunk_size=-1)
true_params = np.array(list(model.get_params().values()))
#model.set_x([logit(random.uniform(.001, join_time-.001) / join_time),
model.set_params([
logit(random.uniform(.001, join_time-.001) / join_time),
random.uniform(-1, 1)],
scaled=True)
res = model.optimize(method="trust-ncg", hessp=True)
inferred_params = np.array(list(model.get_params().values()))
assert np.max(np.abs(np.log(true_params / inferred_params))) < .2
def update_Hess(H, new_x, prev_x, new_g, prev_g):
if np.allclose(new_x, prev_x):
return H
s = new_x - prev_x
y = new_g - prev_g
sy = np.dot(s, y)
Bs = np.linalg.solve(H, s)
y_Bs = y - Bs
if np.abs(np.dot(
s, y_Bs)) < 1e-8 * np.linalg.norm(s) * np.linalg.norm(y_Bs):
# skip SR1 update
return H
Hy = np.dot(H, y)
s_Hy = s - Hy
H = H + np.outer(s_Hy, s_Hy) / np.dot(s_Hy, y)
return H
def callback(weights, iter):
if iter % 10 == 0:
print("Iteration {}".format(iter))
print("\tmax of weights: {}".format(np.max(np.abs(weights))))
cur_loss = loss_fun(weights, train_smiles, train_targets)
training_curve.append(cur_loss)
print("\tloss {}".format(cur_loss))
train_preds = undo_norm(pred_fun(weights, train_smiles))
print("\ttrain {}: {}".format(
nll_func_name, nll_func(train_preds, train_raw_targets)))
if validation_smiles is not None:
validation_preds = undo_norm(
pred_fun(weights, validation_smiles))
print("\tvalidation {}: {}".format(
nll_func_name,
nll_func(validation_preds, validation_raw_targets)))
def test_z_update(self):
# our manual update
test_z_update = dp.z_update(mu, mu2, x, e_info_x, e_log_v, e_log_1mv)
# autograd update
get_auto_z_update = grad(dp.e_loglik_full, 6)
auto_z_update = get_auto_z_update(
x, mu, mu2, tau, e_log_v, e_log_1mv, e_z,
e_info_x, e_logdet_info_x, prior_mu,
prior_inv_wishart_scale, kappa, prior_dof, alpha)
log_const = sp.misc.logsumexp(auto_z_update, axis = 1)
auto_z_update = np.exp(auto_z_update - log_const[:, None])
#print(auto_z_update[0:5, :])
#print(test_z_update[0:5, :])
self.assertTrue(\
np.sum(np.abs(auto_z_update - test_z_update)) <= 10**(-8))
def Fundamental(Y):
# initial guess: use peak of FFT to get to within one sample period
T = float(len(Y)) / np.argmax(np.abs(np.fft.fft(Y)[:2000]))
# (not really necessary, but why not) refine
def mag(T):
w = np.arange(0, len(Y)) * 2 * np.pi / T
x = Y * np.exp(1j * w) # dx/dw = j Y exp(jw)
# dx*/dw = -j Y exp(-jw)
# d/dw = x* dx/dw + x dx*/dw
# = Y exp(1j*w)
return np.abs(np.sum(x))
# print T, mag(T)
# print T+0.1, mag(T+0.1)
# print T-0.1, mag(T-0.1)
# print round(T), mag(round(T))
return round(T)
def callback(weights, iter):
if iter % 10 == 0:
print("max of weights", np.max(np.abs(weights)))
train_preds = undo_norm(
pred_fun(weights, train_smiles[:num_print_examples]))
cur_loss = loss_fun(weights, train_smiles[:num_print_examples],
train_targets[:num_print_examples])
training_curve.append(cur_loss)
print("Iteration", iter, "loss", cur_loss,\
"train RMSE", rmse(train_preds, train_raw_targets[:num_print_examples]),)
if validation_smiles is not None:
validation_preds = undo_norm(
pred_fun(weights, validation_smiles))
print(
"Validation RMSE",
iter,
":",
rmse(validation_preds, validation_raw_targets),
)
def erf(x):
# constants
a1 = 0.254829592
a2 = -0.284496736
a3 = 1.421413741
a4 = -1.453152027
a5 = 1.061405429
p = 0.3275911
# Save the sign of x
sign = np.sign(x)
x = np.abs(x)
# A&S formula 7.1.26
t = 1.0 / (1.0 + p * x)
y = 1.0 - ((((
(a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * np.exp(-x**2)
return sign * y
def variational_objective(params, t):
"""Provides a stochastic estimate of the variational lower bound."""
W, U, B = unpack_params(params)
z0 = np.random.multivariate_normal(np.zeros(dim_z), np.eye(dim_z),
num_samples)
z_prev = z0
sum_log_det_jacob = 0.
for k in range(K):
w, u, b = W[k], U[k], B[k]
u_hat = (m(np.dot(w, u)) - np.dot(w, u)) * (w /
np.linalg.norm(w)) + u
affine = np.outer(h_prime(np.matmul(z_prev, w) + b), w)
sum_log_det_jacob += np.log(eps + np.abs(1 + np.matmul(affine, u)))
z_prev = z_prev + np.outer(h(np.matmul(z_prev, w) + b), u_hat)
z_K = z_prev
log_q_K = -0.5 * np.sum(np.log(2 * np.pi) + z0**2,
1) - sum_log_det_jacob
log_p = np.log(eps + target(z_K))
return np.mean(log_q_K - log_p)
def propagate(
co2dict,
uncertainties_into,
uncertainties_from,
totals=None,
equilibria_in=None,
equilibria_out=None,
dx=1e-6,
dx_scaling="median",
dx_func=None,
):
"""Propagate uncertainties from requested inputs to outputs."""
co2derivs = forward(
co2dict,
uncertainties_into,
uncertainties_from,
totals=totals,
equilibria_in=equilibria_in,
equilibria_out=equilibria_out,
dx=dx,
dx_scaling=dx_scaling,
dx_func=dx_func,
)[0]
npts = np.shape(co2dict["PAR1"])
uncertainties_from = engine.condition(uncertainties_from, npts=npts)[0]
components = {
u_into: {
u_from: np.abs(co2derivs[u_into][u_from]) * v_from
for u_from, v_from in uncertainties_from.items()
}
for u_into in uncertainties_into
}
uncertainties = {
u_into: np.sqrt(
np.sum(
np.array([
component for component in components[u_into].values()
])**2,
axis=0,
))
for u_into in uncertainties_into
}
return uncertainties, components
def _rf_epg(alpha, phi):
if (np.abs(alpha) > 2 * np.pi):
warnings.warn("alpha should be in radians", warnings.UserWarning)
a = np.power(np.cos(alpha / 2.), 2)
b = np.exp(2 * 1j * phi) * np.power(np.sin(alpha / 2.), 2)
c = -1j * np.exp(1j * phi) * np.sin(alpha)
d = np.exp(-2j * phi) * np.power(np.sin(alpha / 2.), 2)
e = np.power(np.cos(alpha / 2.), 2)
f = 1j * np.exp(-1j * phi) * np.sin(alpha)
g = -1j / 2. * np.exp(-1j * phi) * np.sin(alpha)
h = 1j / 2 * np.exp(1j * phi) * np.sin(alpha)
i = np.cos(alpha)
R = np.array([[a, b, c], [d, e, f], [g, h, i]])
return R
def DiederichOpper_II(N, patterns, weights, biases, sc, lr, tol):
'''
rule II described in Diederich and Opper in (1987) Learning of Correlated Patterns in Spin-Glass Networks by Local Learning Rules
'''
for i in range(N): # for each neuron independently
for j in range(patterns.shape[0]):
pattern = np.array(deepcopy(patterns[j].reshape(1, N))).squeeze()
h_i = (weights[i, :] @ pattern.T + biases[i])
# if the new pattern is not already stable with margin 1
while (np.abs(1 - h_i * pattern[i])) > tol:
weights[i, :] = deepcopy(weights[i, :] +
lr * pattern[i] * pattern *
(1 - h_i * pattern[i]))
biases[i] = deepcopy(biases[i] +
lr * pattern[i]) * (1 - h_i * pattern[i])
if sc == True:
weights[i, i] = 0
h_i = (weights[i, :] @ pattern.T + biases[i])
return weights, biases
def ess_compute_Z(diagonal, num_peds, robot_mu_x, robot_mu_y, \
ped_mu_x, ped_mu_y, cov_robot_x, cov_robot_y, \
inv_cov_robot_x, inv_cov_robot_y, cov_ped_x, cov_ped_y, \
inv_cov_ped_x, inv_cov_ped_y, \
one_over_cov_sum_x, one_over_cov_sum_y, normalize):
delta0 = [0. for _ in range(num_peds)]
norm_delta0 = [0. for _ in range(num_peds)]
norm_delta0_normalized = [0. for _ in range(num_peds)]
T = np.size(robot_mu_x)
# for var in range(np.size(var_x_ess)):
for ped in range(num_peds):
# if normalize == True:
# normalize_x = np.multiply(np.power(2*np.pi,-0.5), one_over_std_sum_x)
# normalize_y = np.multiply(np.power(2*np.pi,-0.5), one_over_std_sum_y)
# else:
normalize_x = 1.
normalize_y = 1.
vel_x = robot_mu_x - ped_mu_x[ped]
vel_y = robot_mu_y - ped_mu_y[ped]
vel_x_2 = np.power(vel_x, 2)
vel_y_2 = np.power(vel_y, 2)
one_over_var_sum_x = np.diag(one_over_cov_sum_x[ped])
one_over_var_sum_y = np.diag(one_over_cov_sum_y[ped])
quad_x = np.multiply(one_over_var_sum_x, vel_x_2)
quad_y = np.multiply(one_over_var_sum_y, vel_y_2)
Z_x = np.multiply(normalize_x, np.exp(-0.5 * quad_x))
Z_y = np.multiply(normalize_y, np.exp(-0.5 * quad_y))
Z = np.multiply(Z_x, Z_y)
norm_delta0[ped] = np.abs(np.sum(np.log1p(-Z)))
norm_delta0_normalized = norm_delta0 / (np.sum(norm_delta0))
ess = 1. / np.sum(np.power(norm_delta0_normalized, 2))
ess = np.int(ess)
top_Z_indices = np.argsort(norm_delta0_normalized)[::-1]
return ess, top_Z_indices
def __init__(self, means, variances, pmix=None):
"""
means: a k x d 2d array specifying the means.
variances: a one-dimensional length-k array of variances
pmix: a one-dimensional length-k array of mixture weights. Sum to one.
"""
k, d = means.shape
if k != len(variances):
raise ValueError('Number of components in means and variances do not match.')
if pmix is None:
pmix = old_div(np.ones(k),float(k))
if np.abs(np.sum(pmix) - 1) > 1e-8:
raise ValueError('Mixture weights do not sum to 1.')
self.pmix = pmix
self.means = means
self.variances = variances
def compute_f_fprime_t_avg_12_(W1,W2,perturbation,max_dist=1,burn_in=0.5): # max dist added 10/14/20
#Wmx,Wmy,Wsx,Wsy,s02,K,kappa,T,XX,XXp,Eta,Xi,h1,h2 = parse_W(W)
Wmx,Wmy,Wsx,Wsy,s02,K,kappa,T,h1,h2,bl,amp = parse_W1(W1)
XX,XXp,Eta,Xi = parse_W2(W2)
fval = compute_f_(Eta,Xi,s02)
fprimeval = compute_fprime_(Eta,Xi,s02)
if share_residuals:
resEta = Eta - u_fn(XX,fval,Wmx,Wmy,K,kappa,T)
resXi = Xi - u_fn(XX,fval,Wsx,Wsy,K,kappa,T)
resEta12 = np.concatenate((resEta,resEta),axis=0)
resXi12 = np.concatenate((resXi,resXi),axis=0)
else:
resEta12 = 0
resXi12 = 0
dHH = np.zeros((nN,nQ*nS*nT))
dHH[:,np.arange(2,nQ*nS*nT,nQ)] = 1
dHH = np.concatenate((dHH*h1,dHH*h2),axis=0)
YY = fval + perturbation
YYp = fprimeval
XX12 = np.concatenate((XX,XX),axis=0)
YY12 = np.concatenate((YY,YY),axis=0)
YYp12 = np.concatenate((YYp,YYp),axis=0)
YYmean = np.zeros_like(YY12)
YYprimemean = np.zeros_like(YY12)
def dYYdt(YY,Eta1,Xi1):
return -YY + compute_f_(Eta1,Xi1,s02)
def dYYpdt(YYp,Eta1,Xi1):
return -YYp + compute_fprime_(Eta1,Xi1,s02)
for t in range(niter):
if np.mean(np.abs(YY-fval)) < max_dist:
Eta121 = resEta12 + u_fn(XX12,YY12,Wmx,Wmy,K,kappa,T) + dHH
Xi121 = resXi12 + u_fn(XX12,YY12,Wmx,Wmy,K,kappa,T)
YY12 = YY12 + dt*dYYdt(YY12,Eta121,Xi121)
YYp12 = YYp12 + dt*dYYpdt(YYp12,Eta121,Xi121)
elif np.remainder(t,500)==0:
print('unstable fixed point?')
if t>niter*burn_in:
YYmean = YYmean + 1/niter/burn_in*YY12
YYprimemean = YYprimemean + 1/niter/burn_in*YYp12
#YYmean = YYmean + np.tile(bl,nS*nT)[np.newaxis,:]
return YYmean,YYprimemean
def calc_loss_wrt_parameter_dict(self, param_dict, data_tuple):
''' Compute loss at given parameters
Args
----
param_dict : dict
Keys are string names of parameters
Values are *numpy arrays* of parameter values
Returns
-------
loss : float scalar
'''
# TODO compute loss
y_N = data_tuple[2]
yhat_N = self.predict(data_tuple[0], data_tuple[1], **param_dict)
#loss_total = ag_np.sum(ag_np.square(yhat_N - y_N))
loss_total = ag_np.sum(ag_np.abs(yhat_N - y_N))
return loss_total
def theta_sample(alpha, batch, lr=.2, EM_iter=EM_iter, SGD_iter=SGD_iter):
"""
Returns mc random samples of the posterior mean estimator, for doing Monte Carlo approx.
This uses EM algorithm with _EM_iter_ iterations to converge to approx posterior.
The output is an matrix of _d_ sampled means, times _mc_ samples. Shape = (d, mc)
lr must be maximum 0.25 otherwise explosion occurs with reparam
"""
if model_name == "jaakkola":
lv1 = np.abs(np.random.normal(0, 1, batch[1].shape)) # POSITIVE initial value of lambda(v) before maximization, same shape as Y
for i in range(EM_iter): # EM algorithm updating (mu_P,S_P) and v in turn.
S_P1 = S_P(alpha, batch, lv1) # Cov matrix (d x d)
mu_P1 = mu_P(alpha, batch, S_P1) # Mean vector (d x 1)
lv1 = lambda_v(batch, mu_P1, S_P1) # lambda(v) vector (n x 1)
return np.random.multivariate_normal(mu_P1.reshape(-1), S_P1, mc).T
if model_name == "SVI": # Prior MUST BE N(0,I)
mu_P1 = np.copy(mu_0) # init posterior = prior N(0,I)
rho_P1 = reparam_bwd(sigma_0) # init posterior = prior N(0,I)
gradients_mu = np.empty((d,SGD_iter-1)) # init plotting
for j in range(1, SGD_iter):
epsilon = np.random.multivariate_normal(np.zeros(d), np.identity(d)).reshape(-1, 1) # generate noise ~ N(0,I)
theta = np.nan_to_num( mu_P1 + reparam_fwd(rho_P1.reshape(-1,1)) * epsilon ) # nan to num is used to fix Autograd bug with sqrt(0)
df_dthetha1 = np.nan_to_num( df_dtheta(theta, alpha, batch, mu_P1, rho_P1) )
grad_mu_P = np.nan_to_num( df_dthetha1 + df_dmu_P(theta, alpha, batch, mu_P1, rho_P1) )
grad_rho_P = np.nan_to_num( df_dthetha1 * (epsilon/(1 + np.exp(-rho_P1.reshape(-1,1)))) + df_drho_P(theta, alpha, batch, mu_P1, rho_P1).reshape(-1,1) )
mu_P1 -= lr/np.sqrt(j) * grad_mu_P # gradient descent
rho_P1 -= lr/np.sqrt(j) * np.squeeze(grad_rho_P)
# Plotting the SGD
gradients_mu[:,j-1] = np.squeeze(mu_P1)
if(False and alpha==0):
plt.plot(gradients_mu.T)
plt.xlabel("iteration")
plt.ylabel("mu_P's values")
plt.savefig("../plots/SGD_SVI.png")
plt.show()
plt.clf()
return np.random.multivariate_normal(mu_P1.reshape(-1), np.diag(reparam_fwd(rho_P1)**2), mc).T
def get_log_pc(self, v):
logit_v = np.log(v) - np.log(1 - v)
epsilon = self.epsilon_param
if np.abs(epsilon) < 1e-8:
return self.get_log_p0(v)
if self.gustafson_style:
log_epsilon = np.log(epsilon)
return \
self.get_log_p0(v) + \
self.log_phi(logit_v) + \
log_epsilon - \
self.log_norm_pc
else:
# assert epsilon <= 1
return \
self.get_log_p0(v) + \
epsilon * self.log_phi(logit_v) - \
self.log_norm_pc
def plot_reverse_animation(self, fig, ax, x, y, z, threshold=1e-5, MAX=1000):
frames = 48
current_angle = self.get_angle()
# Update using reverse kinematic
self.update_reverse_kinematic(x, y, z, threshold, MAX)
new_angle = self.get_angle()
# Angle change by frame
diff_angle = new_angle - current_angle
for i in range(len(diff_angle)):
if np.abs(diff_angle[i]) > np.pi:
if diff_angle[i] > 0:
diff_angle[i] = diff_angle[i] - PI2
else:
diff_angle[i] = PI2 + diff_angle[i]
change_angle = diff_angle / frames
lines = self.plot_arm(ax, current_angle)
def get_plot(i):
plot_angle = current_angle + i * change_angle
first = self.get_arm_pos(1, *plot_angle)
second = self.get_arm_pos(2, *plot_angle)
third = self.get_arm_pos(3, *plot_angle)
lines[0].set_data(*list(zip([0,0], first[:2])))
lines[0].set_3d_properties([0, first[2]])
lines[1].set_data(*list(zip(first[:2], second[:2])))
lines[1].set_3d_properties([first[2], second[2]])
lines[2].set_data(*list(zip(second[:2], third[:2])))
lines[2].set_3d_properties([second[2], third[2]])
return lines
anim = animation.FuncAnimation(fig, get_plot, frames=48, interval=100, blit=True)
return anim
def variational_log_density(params, samples):
'''
samples: [n_samples, D]
u: [D,1]
w: [D,1]
b: [1]
Returns: [num_samples]
'''
n_samples = len(samples)
mean = params[0]
log_std = params[1]
norm_flow_params = params[2]
z_k, all_zs = normalizing_flows(samples, norm_flow_params)
logp_zk = logprob(z_k)
logp_zk = np.reshape(logp_zk, [n_samples, 1])
logq_z0 = diag_gaussian_log_density(samples, mean, log_std)
logq_z0 = np.reshape(logq_z0, [n_samples, 1])
sum_nf = np.zeros([n_samples,1])
for params_k in range(len(norm_flow_params)):
u = norm_flow_params[params_k][0]
w = norm_flow_params[params_k][1]
b = norm_flow_params[params_k][2]
# Appendix equations
m_x = -1. + np.log(1.+np.exp(np.dot(w.T,u)))
u_k = u + (m_x - np.dot(w.T,u)) * (w/np.linalg.norm(w))
# u_k = u
# [n_samples, D]
phi = np.dot((1.-np.tanh(np.dot(all_zs[params_k],w)+b)**2), w.T)
# [n_samples, 1]
sum_nf = np.log(np.abs(1+np.dot(phi, u_k)))
sum_nf += sum_nf
# return logq_z0 - sum_nf
log_qz = np.reshape(logq_z0 - sum_nf, [n_samples])
return log_qz
def print_wf_values(theta1=0.0, theta2=0.0, use_j=False, B=0.0):
wf = Wavefunction(use_jastrow=use_j)
# Adjust numpy output so arrays are printed with higher precision
float_formatter = "{:.15g}".format
np.set_printoptions(formatter={'float_kind':float_formatter})
if use_j:
VP = np.array([theta1, theta2, B])
print("Values for theta = ",theta1,theta2," and jastrow B = ",B)
else:
VP = np.array([theta1, theta2])
print("Values for theta = ",theta1,theta2," and no jastrow")
r1 = np.array([1.0, 2.0, 3.0])
r2 = np.array([0.0, 1.1, 2.2])
psi_val = wf.psi(r1, r2, VP)
print(" wf = ",psi_val," log wf = ",np.log(np.abs(psi_val)))
g0 = wf.grad0(r1, r2, VP)/psi_val
print(" grad/psi for particle 0 = ",g0[0],g0[1],g0[2])
# Using the laplacian of log psi to match internal QMCPACK values
lap_0 = wf.lap0(r1, r2, VP)
print(" laplacian of log psi for particle 0 = ",lap_0)
lap_1 = wf.lap1(r1, r2, VP)
print(" laplacian for log psi particle 1 = ",lap_1)
eloc = wf.local_energy(r1, r2, VP)
print(" local energy = ",eloc)
dp = wf.dpsi(r1, r2, VP)
print(" parameter derivative of log psi = ",dp / psi_val)
deloc = wf.dlocal_energy(r1, r2, VP)
print(" parameter derivative of local energy = ",deloc)
print("")
def run_cavi(self, tau, nu, phi_mu, phi_var, max_iter=200, tol=1e-6):
params = packing.flatten_params(tau, nu, phi_mu, phi_var)
self.trace.reset()
diff = np.float('inf')
while diff > tol and self.trace.stepnum < max_iter:
self.cavi_updates(tau, nu, phi_mu, phi_var)
new_params = packing.flatten_params(tau, nu, phi_mu, phi_var)
diff = np.max(np.abs(new_params - params))
self.trace.update(params, diff)
if not np.isfinite(diff):
print('Error: non-finite parameter difference.')
break
params = new_params
if self.trace.stepnum >= max_iter:
print('Warning: CAVI reached max_iter.')
print('Done with CAVI.')
return tau, nu, phi_mu, phi_var
def predict(self, test_x, is_diag=1):
sn2 = np.exp(self.theta[0])
hyp = self.theta[1:]
tmp = self.kernel(test_x, self.train_x, hyp)
py = np.dot(tmp, self.alpha)
'''
ps2 = sn2 + self.kernel(test_x, test_x, hyp) - np.dot(tmp, chol_inv(self.L, tmp.T))
if is_diag:
ps2 =np.diag(ps2)
'''
tmp1 = chol_inv(self.L, tmp.T)
# ps2 = -np.dot(tmp, chol_inv(self.L, tmp.T))
if is_diag:
ps2 = self.for_diag + sn2 - (tmp * tmp1.T).sum(axis=1)
else:
ps2 = sn2 - np.dot(tmp, tmp1) + self.kernel(test_x, test_x, hyp)
ps2 = np.abs(ps2)
py = py * self.std + self.mean
ps2 = ps2 * (self.std**2)
return py, ps2
def __init__(self, n, penalty='huber', alpha=1.0):
assert (alpha > 0.0)
self.alpha = alpha
self.alpha_sq = alpha ** 2
self.penalty = penalty.lower()
if (self.penalty == 'quadratic'):
self.phi = lambda z: 0.5 * np.power(z, 2.0)
elif (self.penalty == 'pseudo-huber'):
self.phi = lambda z: self.alpha_sq * (np.sqrt(1.0 + np.power(z, 2.0) / self.alpha_sq) - 1.0)
elif (self.penalty == 'huber'):
self.phi = lambda z: np.where(np.abs(z) <= alpha, 0.5 * np.power(z, 2.0), alpha * np.abs(z) - 0.5 * self.alpha_sq)
elif (self.penalty == 'welsch'):
self.phi = lambda z: 1.0 - np.exp(-0.5 * np.power(z, 2.0) / self.alpha_sq)
elif (self.penalty == 'trunc-quad'):
self.phi = lambda z: np.minimum(0.5 * np.power(z, 2.0), 0.5 * self.alpha_sq)
else:
assert False, "unrecognized penalty function {}".format(penalty)
super().__init__(n, 1) # make sure node is properly constructed
self.eps = 1.0e-4 # relax tolerance on optimality test
def predict(self, test_x, is_diag=1):
output_scale = np.exp(self.theta[0])
sigma2_tag = np.exp(self.theta[self.dim+2])
C = self.kernel(self.src_x, self.tag_x, self.theta)
L_C = np.linalg.cholesky(C)
alpha_C = chol_inv(L_C, self.train_y.T)
k_star_s = self.kernel2(test_x, self.src_x, self.theta)
k_star_t = self.kernel1(test_x, self.tag_x, self.theta)
k_star = np.hstack((k_star_s, k_star_t))
py = np.dot(k_star, alpha_C)
Cvks = chol_inv(L_C, k_star.T)
if is_diag:
ps2 = output_scale + sigma2_tag - (k_star * Cvks.T).sum(axis=1)
else:
ps2 = self.kernel1(test_x, test_x, self.theta) + sigma2_tag - np.dot(k_star, Cvks)
ps2 = np.abs(ps2)
py = py * self.std + self.mean
ps2 = ps2 * (self.std**2)
return py, ps2
def lanczos(dx, a=3):
"""Lanczos kernel
Parameters
----------
dx: float
amount to shift image
a: int
Lanczos window size parameter
Returns
-------
result: array-like
1D Lanczos kernel
"""
if np.abs(dx) > 1:
raise ValueError("The fractional shift dx must be between -1 and 1")
window = np.arange(-a + 1, a + 1) + np.floor(dx)
y = np.sinc(dx - window) * np.sinc((dx - window) / a)
return y, window.astype(int)
def forwardFilter( self, knownLatentStates=None ):
if( knownLatentStates is not None ):
assert np.abs( knownLatentStates - knownLatentStates.astype( int ) ).sum() == 0.0
knownLatentStates = knownLatentStates.astype( int )
if( knownLatentStates.size == 0 ):
self.chainCuts = None
else:
# Assert that knownLatentStates is sorted
assert np.any( np.diff( knownLatentStates[ :, 0 ] ) <= 0 ) == False
# Mark that we are cutting the markov chain at these indices
self.chainCuts = knownLatentStates
else:
self.chainCuts = None
return super( CategoricalHMM, self ).forwardFilter()
