def image_bbox(params, img):
img_ymax, img_xmax = img.nelec.shape
px, py = img.equa2pixel(params.u)
xlim = (np.max([0, int(np.floor(px - pixel_radius))]),
np.min([img_xmax, int(np.ceil(px + pixel_radius))]))
ylim = (np.max([0, int(np.floor(py - pixel_radius))]),
np.min([img_ymax, int(np.ceil(py + pixel_radius))]))
return xlim, ylim
def get_bounding_box(params, img):
if params.is_star():
bound = img.R
elif params.is_galaxy():
bound = gal_funs.gen_galaxy_psf_image_bound(params, img)
else:
raise "source type unknown"
px, py = img.equa2pixel(params.u)
xlim = (np.max([0, np.floor(px - bound)]),
np.min([img.nelec.shape[1], np.ceil(px + bound)]))
ylim = (np.max([0, np.floor(py - bound)]),
np.min([img.nelec.shape[0], np.ceil(py + bound)]))
return xlim, ylim
def plot(self, name, plot_std=False):
if self.X.shape[1] > 1:
raise Exception('Dimension of X should be 1 for this method...')
x = np.linspace(np.min(self.X), np.max(self.X), 100).reshape(-1, 1)
self.optimize(restart=2)
self.likelihood(self.params)
mean, std = self.inference(x, return_std=True)
plt.plot(x, mean, "--", label='GPR-' + str(name), color='deepskyblue')
plt.scatter(self.X,
self.y,
label='GPR-Train' + str(name),
color='deepskyblue')
if plot_std is True:
plt.fill_between(x.ravel(),
mean.ravel() + 2. * std,
mean.ravel() - 2. * std,
alpha=0.2,
color='deepskyblue')
plt.fill_between(x.ravel(),
mean.ravel() + 1. * std,
mean.ravel() - 1. * std,
alpha=0.3,
color='deepskyblue')
plt.xlabel('$x$')
plt.ylabel('$y$')
plt.legend()
def plot_cost_history(self, ax, history, start):
# plotting colors
colors = ['k']
# plot cost function history
ax.plot(np.arange(start, len(history), 1),
history[start:],
linewidth=3,
color='k')
# clean up panel / axes labels
xlabel = 'step $k$'
ylabel = r'$g\left(\mathbf{w}^k\right)$'
ax.set_xlabel(xlabel, fontsize=14)
ax.set_ylabel(ylabel, fontsize=14, rotation=0, labelpad=25)
title = 'cost history'
ax.set_title(title, fontsize=18)
# plotting limits
xmin = 0
xmax = len(history)
xgap = xmax * 0.05
xmin -= xgap
xmax += xgap
ymin = np.min(history)
ymax = np.max(history)
ygap = ymax * 0.05
ymin -= ygap
ymax += ygap
ax.set_xlim([xmin, xmax])
ax.set_ylim([ymin, ymax])
def geom(self, x, geom_ord=[0], k=100):
loglik = None
agrad = None
HessApply = None
eigs = None
# get log-likelihood
if any(s >= 0 for s in geom_ord):
loglik = -self.cost(x)
# get gradient
if any(s >= 1 for s in geom_ord):
g = np.zeros_like(x)
agrad = -self.grad(x, g)
# get Hessian Apply
if any(s >= 1.5 for s in geom_ord):
HessApply = None
# get estimated eigen-decomposition for the Hessian (or Gauss-Newton)
if any(s > 1 for s in geom_ord):
# eigs = (np.array([1., 0.1]), np.array([np.ones_like(x),-np.ones_like(x)]))
# eigs = (np.ones(1), np.ones_like(x))
eigs = self.eigdecomp(x, k=np.min([self.dimension, k]))
return loglik, agrad, HessApply, eigs
def reparam_gradient(alpha,m,x,K,alphaz,corr=True,B=0):
gradient = np.zeros((alpha.shape[0],2))
if B == 0:
assert np.min(alpha)>= 1.,"Needs alpha boost"
lmbda = npr.gamma(alpha,1.)
lmbda[lmbda < 1e-300] = 1e-300
zw = m*lmbda/alpha
epsilon = calc_epsilon(lmbda, alpha)
h_val = gamma_h(epsilon, alpha)
h_der = gamma_grad_h(epsilon, alpha)
logp_der = grad_logp(zw, K, x, alphaz)
gradient[:,0] = logp_der*m*(alpha*h_der-h_val)/alpha**2
gradient[:,1] = logp_der*h_val/alpha
gradient += grad_entropy(alpha,m)
if corr:
gradient[:,0] += logp(zw, K, x, alphaz)*gamma_correction(epsilon, alpha)
else:
lmbda = npr.gamma(alpha+B,1.)
lmbda[lmbda < 1e-5] = 1e-5
u = npr.rand(alpha.shape[0],B)
epsilon = calc_epsilon(lmbda, alpha+B)
h_val = gamma_h_boosted(epsilon,u,alpha)
h_der = gamma_grad_h_boosted(epsilon,u,alpha)
zw = h_val*m/alpha
zw[zw < 1e-5] = 1e-5
logp_der = grad_logp(zw, K, x, alphaz)
gradient[:,0] = logp_der*m*(alpha*h_der-h_val)/alpha**2
gradient[:,1] = logp_der*h_val/alpha
gradient += grad_entropy(alpha,m)
if corr:
gradient[:,0] += logp(zw, K, x, alphaz)*gamma_correction(epsilon, alpha+B)
return gradient
def plot_images(images,
ax,
ims_per_row=5,
padding=5,
digit_dimensions=(28, 28),
cmap=matplotlib.cm.binary,
vmin=None,
vmax=None):
"""Images should be a (N_images x pixels) matrix."""
N_images = images.shape[0]
N_rows = np.ceil(float(N_images) / ims_per_row)
pad_value = np.min(images.ravel())
concat_images = np.full(
((digit_dimensions[0] + padding) * N_rows + padding,
(digit_dimensions[1] + padding) * ims_per_row + padding), pad_value)
for i in range(N_images):
cur_image = np.reshape(images[i, :], digit_dimensions)
row_ix = i // ims_per_row
col_ix = i % ims_per_row
row_start = padding + (padding + digit_dimensions[0]) * row_ix
col_start = padding + (padding + digit_dimensions[1]) * col_ix
concat_images[row_start:row_start + digit_dimensions[0],
col_start:col_start + digit_dimensions[1]] = cur_image
cax = ax.matshow(concat_images, cmap=cmap, vmin=vmin, vmax=vmax)
plt.xticks(np.array([]))
plt.yticks(np.array([]))
return cax
def get_sharp_TE_airfoil(self):
# Returns a version of the airfoil with a sharp trailing edge.
upper_original_coors = self.upper_coordinates(
) # Note: includes leading edge point, be careful about duplicates
lower_original_coors = self.lower_coordinates(
) # Note: includes leading edge point, be careful about duplicates
# Find data about the TE
# Get the scale factor
x_mcl = self.mcl_coordinates[:, 0]
x_max = np.max(x_mcl)
x_min = np.min(x_mcl)
scale_factor = (x_mcl - x_min) / (x_max - x_min) # linear contraction
# Do the contraction
upper_minus_mcl_adjusted = self.upper_minus_mcl - self.upper_minus_mcl[
-1, :] * np.expand_dims(scale_factor, 1)
# Recreate coordinates
upper_coordinates_adjusted = np.flipud(self.mcl_coordinates +
upper_minus_mcl_adjusted)
lower_coordinates_adjusted = self.mcl_coordinates - upper_minus_mcl_adjusted
coordinates = np.vstack(
(upper_coordinates_adjusted[:-1, :], lower_coordinates_adjusted))
# Make a new airfoil with the coordinates
name = self.name + ", with sharp TE"
new_airfoil = Airfoil(name=name,
coordinates=coordinates,
repanel=False)
return new_airfoil
def hmc(self, position_current, momentum_current):
### Refresh momentum
momentum_current = self.sample_momentum(1)
### Simulate Hamiltonian dynamics using Leap Frog
position_proposal, momentum_proposal = self.leap_frog(position_current, momentum_current)
# compute total energy in current position and proposal position
current_total_energy = self.total_energy(position_current, momentum_current)
proposal_total_energy = self.total_energy(position_proposal, momentum_proposal)
### Output for diganostic mode
if self.params['diagnostic_mode']:
print('potential energy change:',
self.potential_energy(position_current),
self.potential_energy(position_proposal))
print('kinetic energy change:',
self.kinetic_energy(momentum_current),
self.kinetic_energy(momentum_proposal))
print('total enregy change:',
current_total_energy,
proposal_total_energy)
print('\n\n')
### Metropolis Hastings Step
# comute accept probability
accept_prob = np.min([1, np.exp(current_total_energy - proposal_total_energy)])
# accept proposal with accept probability
if self.random.rand() < accept_prob:
self.accepts += 1.
position_current = np.copy(position_proposal)
momentum_current = momentum_proposal
return position_current, momentum_current
def get_camber_at_chord_fraction_legacy(self, chord_fraction):
# Returns the (interpolated) camber at a given location(s). The location is specified by the chord fraction, as measured from the leading edge. Camber is nondimensionalized by chord (i.e. this function returns camber/c at a given x/c).
chord = np.max(self.coordinates[:, 0]) - np.min(
self.coordinates[:, 0]
) # This should always be 1, but this is just coded for robustness.
x = chord_fraction * chord + min(self.coordinates[:, 0])
upperCoors = self.upper_coordinates()
lowerCoors = self.lower_coordinates()
y_upper_func = sp_interp.interp1d(x=upperCoors[:, 0],
y=upperCoors[:, 1],
copy=False,
fill_value='extrapolate')
y_lower_func = sp_interp.interp1d(x=lowerCoors[:, 0],
y=lowerCoors[:, 1],
copy=False,
fill_value='extrapolate')
y_upper = y_upper_func(x)
y_lower = y_lower_func(x)
camber = (y_upper + y_lower) / 2
return camber
def variational_objective(params, t, num_samples, beta=1.):
"""Provides a stochastic estimate of the variational lower bound."""
# 1. draw samples from the variational posterior, eps ~ N(0,I)
zs, ldet_sums = draw_variational_samples(params, num_samples)
# 1.5 negative entropy of z0 --- likely we need this for KL though
# not needed for optimization
# 2. compute expected value of the sum of jacobian terms
E_ldet_sum = np.mean(ldet_sums)
# 3. compute data term
lls = logprob(zs, t)
E_logprob = np.mean(lls)
if debug_print:
print "entropy term: ", E_ldet_sum
print "data term : ", E_logprob, " (+/- ", np.std(
lls), ")", " min = ", np.min(lls)
# return lower bound
beta = 1. if t >= len(beta_schedule) else beta_schedule[t]
lower_bound = beta * E_logprob + E_ldet_sum
return -lower_bound
def line_point_dist(lines, ps):
"""
Closest distance of a point to a line segment defined by two points (a, b).
The arguments can also be lists of lines and points, in that case the distance for
each combination is returned, with shape lines.shape[:-2] + ps.shape[:-1].
"""
assert(lines.shape[-2:] == (2, 2))
assert(ps.shape[-1] == 2)
a = lines[...,0,:]
b = lines[...,1,:]
for _ in range(max(len(ps.shape)-1, 1)):
a = np.expand_dims(a, -2)
b = np.expand_dims(b, -2)
# ps = np.expand_dims(ps, 0)
v_hat = (b - a) / np.expand_dims(norm(b - a), -1)
# d_along.shape == (v_hat.shape[0], ps.shape[0])
# i.e. one scalar product for each line-point combination
d_along = np.sum(v_hat*(ps - a), axis=-1)
d_normal = np.abs(cross(v_hat, ps - a))
assert(d_along.shape == d_normal.shape)
d_ends = np.min(np.array([norm(ps-a), norm(ps-b)]), axis=0)
# if p lies along the sides of the line use the normal distance,
# else the distance to one of the ends
mask = (0 <= d_along) & (d_along <= norm(b - a))
return np.where(mask, d_normal, d_ends)
def expected_improvement(x, gaussian_process, evaluated_loss):
""" expected_improvement
Expected improvement acquisition function.
Arguments:
----------
x: array-like, shape = [n_samples, n_hyperparams]
The point for which the expected improvement needs to be computed.
gaussian_process: GaussianProcessRegressor object.
Gaussian process trained on previously evaluated hyperparameters.
evaluated_loss: Numpy array.
Numpy array that contains the values off the loss function for the previously
evaluated hyperparameters.
"""
x_to_predict = x.reshape(1, -1)
mu, sigma = gaussian_process.predict(x_to_predict, return_std=True)
loss_optimum = np.min(evaluated_loss)
# In case sigma equals zero
with np.errstate(divide='ignore'):
Z = (mu - loss_optimum) / sigma
expected_improvement = (
mu - loss_optimum) * norm.cdf(Z) + sigma * norm.pdf(Z)
expected_improvement[sigma == 0.0] == 0.0
return expected_improvement
def random_eval_experiment():
'''
Experiment illutrating how quickly global random evaluation will fail as a method of optimization. Output is minimum value attained by random sampling over the cube [-1,1] x [-1,1] x... [-1,1] evaluating simple quadratic for 100, 1000, or 10000 times. The dimension is increased from 1 to 100 and the minimum plotted for each dimension.
'''
# define symmetric quadratic N-dimensional
g = lambda w: np.dot(w.T, w)
# loop over dimensions, sample points, evaluate
mean_evals = []
big_dim = 100
num_pts = 10000
pt_stops = [100, 1000, 10000]
for dim in range(big_dim):
dim_eval = []
m_eval = []
for pt in range(num_pts):
# generate random point using uniform
r = 2 * np.random.rand(dim + 1) - 1
e = g(r)
dim_eval.append(e)
# record mean and std of so many pts
if (pt + 1) in pt_stops:
m_eval.append(np.min(dim_eval))
mean_evals.append(m_eval)
# convert to array for easy access
mean_evals_global = np.asarray(mean_evals)
fig = plt.figure(figsize=(6, 3))
# create subplot with 3 panels, plot input function in center plot
gs = gridspec.GridSpec(1, 1, width_ratios=[1])
fig.subplots_adjust(wspace=0.5, hspace=0.01)
# plot input function
ax = plt.subplot(gs[0])
for k in range(len(pt_stops)):
mean_evals = mean_evals_global[:, k]
# scatter plot mean value
ax.plot(np.arange(big_dim) + 1, mean_evals)
# clean up plot - label axes, etc.,
ax.set_xlabel('dimension of input')
ax.set_ylabel('funciton value')
# draw legend
t = [str(p) for p in pt_stops]
ax.legend(t, bbox_to_anchor=(1, 0.5))
# draw horizontal axis
ax.plot(np.arange(big_dim) + 1,
np.arange(big_dim) * 0,
linewidth=1,
linestyle='--',
color='k')
plt.show()
def _parameter_initialiser(self, x, c=None, n=None, offset=False):
log_x = np.log(x)
log_x[np.isnan(log_x)] = 0
gumb = para.Gumbel.fit(log_x, c, n, how='MLE')
if not gumb.res.success:
gumb = para.Gumbel.fit(log_x, c, n, how='MPP')
mu, sigma = gumb.params
alpha, beta = np.exp(mu), 1. / sigma
if (np.isinf(alpha) | np.isnan(alpha)):
alpha = np.median(x)
if (np.isinf(beta) | np.isnan(beta)):
beta = 1.
if offset:
gamma = np.min(x) - (np.max(x) - np.min(x)) / 10.
return gamma, alpha, beta, 1.
else:
return alpha, beta, 1.
def show_2d_classifier(self, ax, w_best, run, **kwargs):
cost = run.cost
predict = run.model
feat = run.feature_transforms
normalizer = run.normalizer
### create surface and boundary plot ###
xmin1 = np.min(copy.deepcopy(self.x[:, 0]))
xmax1 = np.max(copy.deepcopy(self.x[:, 0]))
xgap1 = (xmax1 - xmin1) * 0.05
xmin1 -= xgap1
xmax1 += xgap1
xmin2 = np.min(copy.deepcopy(self.x[:, 1]))
xmax2 = np.max(copy.deepcopy(self.x[:, 1]))
xgap2 = (xmax2 - xmin2) * 0.05
xmin2 -= xgap2
xmax2 += xgap2
# plot boundary for 2d plot
r1 = np.linspace(xmin1, xmax1, 500)
r2 = np.linspace(xmin2, xmax2, 500)
s, t = np.meshgrid(r1, r2)
s = np.reshape(s, (np.size(s), 1))
t = np.reshape(t, (np.size(t), 1))
h = np.concatenate((s, t), axis=1)
z = predict(normalizer(h.T), w_best)
z = np.sign(z)
# reshape it
s.shape = (np.size(r1), np.size(r2))
t.shape = (np.size(r1), np.size(r2))
z.shape = (np.size(r1), np.size(r2))
#### plot contour, color regions ####
ax.contour(s, t, z, colors='k', linewidths=2.5, levels=[0], zorder=2)
ax.contourf(s,
t,
z,
colors=[self.colors[1], self.colors[0]],
alpha=0.15,
levels=range(-1, 2))
# cleanup panel
ax.set_xlim([xmin1, xmax1])
ax.set_ylim([xmin2, xmax2])
def timing_error(bubble_position, piezo_timings):
"""Calculate the mean squared error of the expected timings for a certain bubble positions versus those actually observed"""
# Get the expected overall times of flight to the piezos for this bubble
times_of_flight = expected_times_of_flight(bubble_position)
# Subtract the minimum time of flight from all of them, so that the first signal is at time 0
expected_timings = times_of_flight - np.min(times_of_flight)
# Return the mean squared error between the actually observed piezo timings and the timings that would be expected for this bubble
return np.linalg.norm(expected_timings - piezo_timings)
def get_starlet_shape(shape, lvl=None):
""" Get the pad shape for a starlet transform
"""
#Number of levels for the Starlet decomposition
lvl_max = np.int(np.log2(np.min(shape[-2:])))
if (lvl is None) or lvl > lvl_max:
lvl = lvl_max
return int(lvl)
def ab_cb(self, x, a, b, N, alpha=0.05):
# Parameter confidence intervals from here:
# https://mathoverflow.net/questions/278675/confidence-intervals-for-the-endpoints-of-the-uniform-distribution
#
sample_range = np.max(x) - np.min(x)
fun = lambda c : self.p(c, N)
c_hat = minimize(fun, 1.).x
return a - c_hat*sample_range, b + c_hat*sample_range
def get_standard_error_matrix(betahat, y, x, w, se_group=None):
"""Return the standard error matrix for the regression estimate betahat.
If se_group is None, compute the ordinary regression standard error.
Otherwise, compute the robust standard errors using the grouping given by
se_group, which is assumed to be integers 0:(num_groups - 1).
Note that se_group must be zero-indexed, and the number of groups is taken
to be the largest index plus one. (This behavior is implicitly assumed in
group_sum.)
With the se_group option, no finite-sample bias adjustment is applied.
For example, the resulting ses should be equivalent to calling the
R function
sandwich::vcovCL(..., cluster=se_group, type="HC0", cadjust=FALSE)
"""
resid = y - x @ betahat
# For now, I am taking the weights to parameterize a change to the
# objective function rather than a change to the empirical distribution.
# See email from me to Rachael and Tamara on Jan 31, 2020, 2:50 PM
# for more discussion of this subtle point.
if se_group is None:
# I am using num_obs instead of np.sum(w) because w does not
# parameterize the empirical distribution.
num_obs = len(y)
xtx_bar = np.einsum('ni,nj,n->ij', x, x, w) / num_obs
sigma2hat = np.sum(w * (resid ** 2)) / (num_obs - len(betahat))
xtx_inv = np.linalg.inv(xtx_bar)
se2 = sigma2hat * xtx_inv / num_obs
return se2
else:
if len(se_group) != len(y):
raise ValueError("se_group must be the same length as the data.")
#resid = y - x @ betahat
if np.min(se_group) != 0:
raise ValueError('se_group must be zero-indexed ' +
'(its minimum must be zero)')
# Calculate the sample variance of the gradient where each group
# is treated as a single observation.
grad = w[:, None] * resid[:, None] * x
grad_grouped = grouped_sum(grad, se_group)
num_groups = grad_grouped.shape[0]
grad2_mean = np.einsum('gi,gj->ij',
grad_grouped, grad_grouped) / num_groups
grad_mean = np.einsum('gi->i', grad_grouped) / num_groups
grad_cov = grad2_mean - np.outer(grad_mean, grad_mean)
# Weight by the Hessian.
xtx_bar = np.einsum('ni,nj,n->ij', x, x, w) / num_groups
hinv_grad_cov = np.linalg.solve(xtx_bar, grad_cov)
se2 = np.linalg.solve(xtx_bar, hinv_grad_cov.T) / num_groups
return se2
def plot_multiple_sequences(seq1, seq2, seq3):
# initialize figure
fig = plt.figure(figsize=(10, 5))
# create subplot with 3 panels, plot input function in center plot
gs = gridspec.GridSpec(2, 1)
ax1 = plt.subplot(gs[1])
ax2 = plt.subplot(gs[0])
ax1.plot(np.arange(np.size(seq1)), seq1.flatten(), c='k', linewidth=2.5)
ax2.plot(np.arange(np.size(seq2)),
seq2.flatten(),
c='lime',
linewidth=2.5,
label='sequence 1',
zorder=2)
ax2.plot(np.arange(np.size(seq3)),
seq3.flatten(),
c='m',
linewidth=2.5,
label='sequence 2',
zorder=1)
# label axes and title
ax1.set_title('input sequence')
ax1.set_xlabel('step')
ax2.set_title('output sequences')
ax2.set_xlabel('step')
# set viewing limits
s1min = np.min(copy.deepcopy(seq1))
s1max = np.max(copy.deepcopy(seq1))
s1gap = (s1max - s1min) * 0.1
s1min -= s1gap
s1max += s1gap
ax1.set_ylim([s1min, s1max])
s2min = np.min(copy.deepcopy(seq2))
s2max = np.max(copy.deepcopy(seq2))
s2gap = (s2max - s2min) * 0.1
s2min -= s2gap
s2max += s2gap
ax2.legend(loc=1)
plt.show()
def prox_sdss_symmetry(X, step):
"""SDSS/HSC symmetry operator
This function uses the *minimum* of the two
symmetric pixels in the update.
"""
Xs = np.fliplr(np.flipud(X))
X[:] = np.min([X, Xs], axis=0)
return X
def render(self, model):
"""Resample and convolve a model in the observation frame for display only!
Parameters
----------
model: array
The model in some other data frame.
Returns
-------
model_: array
The convolved and resampled `model` in the observation frame.
"""
img = np.zeros(self.frame.shape)
img[:, np.min(self._coord_lr[0]).astype(int):np.max(self._coord_lr[0]).astype(int) + 1,
np.min(self._coord_lr[1]).astype(int):np.max(self._coord_lr[1]).astype(int) + 1] = \
self._render(model)
return img
def add_jitter(kernel, jitter=1e-5):
# Add the jitter
diag_indices = np.diag_indices(np.min(kernel.shape[:2]))
to_add = np.zeros_like(kernel)
to_add[diag_indices] += jitter
kernel = kernel + to_add
return kernel
def to_lane_dist(lanes_col1, lanes_col2):
# use middle point to calcu dist_y?
sample_num = lanes_col1.shape[0]
dist_y = np.min(abs(np.hstack((lanes_col1[:, 1].reshape(sample_num, 1), lanes_col2[:, 1].reshape(sample_num, 1)))), axis=1) # nearest y in each lane
theta = np.arctan((lanes_col1[:, 0] - lanes_col2[:, 0]) / (lanes_col2[:, 1] - lanes_col1[:, 1]))
rho = abs(lanes_col1[:, 0] * np.cos(theta) + lanes_col1[:, 1] * np.sin(theta)) # offset to lane
dist = dist_y + rho
return dist
def __init__(self,
latent_dim,
noise_dim,
model_directory,
latent=None,
full=False,
config_fname='op_conditions.ini'):
self.latent = latent
self.latent_dim = latent_dim
self.noise_dim = noise_dim
if noise_dim == 0:
full = False
self.full = full
if (not full) and (self.latent is None):
self.dim = self.latent_dim
self.bounds = np.array([[0., 1.]])
self.bounds = np.tile(self.bounds, [self.dim, 1])
else:
self.dim = self.latent_dim + self.noise_dim
if self.latent is not None:
assert len(self.latent) == self.latent_dim
latent_bounds = np.vstack((latent - 0.1, latent + 0.1)).T
else:
latent_bounds = np.array([0., 1.])
latent_bounds = np.tile(latent_bounds, [self.latent_dim, 1])
noise_bounds = np.array([-0.5, 0.5])
noise_bounds = np.tile(noise_bounds, [self.noise_dim, 1])
self.bounds = np.vstack((latent_bounds, noise_bounds))
# Expand bounds by 20%
b = self.bounds
r = np.max(b, axis=1) - np.min(b, axis=1)
self.bounds = np.zeros_like(b)
self.bounds[:, 0] = b[:, 0] - 0.2 * r
self.bounds[:, 1] = b[:, 1] + 0.2 * r
self.y = None
self.config_fname = config_fname
self.gan = GAN(self.latent_dim, self.noise_dim, 192, 31, (0., 1.))
self.gan.restore(model_directory)
n_points = self.gan.X_shape[0]
x_synth = self.gan.x_fake_test
x_synth_ = tf.squeeze(x_synth)
self.x_target = tf.placeholder(tf.float32, shape=[n_points, 2])
self.e = tf.reduce_mean(
tf.reduce_sum(tf.square(x_synth_ - self.x_target), axis=1))
if self.full:
self.grad_e = tf.concat(tf.gradients(self.e,
[self.gan.c, self.gan.z]),
axis=1)
else:
self.grad_e = tf.gradients(self.e, self.gan.c)
def adagrad_optimize(n_iters,
objective_and_grad,
init_param,
has_log_norm=False,
window=10,
learning_rate=.01,
epsilon=.1,
learning_rate_end=None):
local_grad_history = []
local_log_norm_history = []
value_history = []
log_norm_history = []
variational_param = init_param.copy()
variational_param_history = []
with tqdm.trange(n_iters) as progress:
try:
schedule = learning_rate_schedule(n_iters, learning_rate,
learning_rate_end)
for i, curr_learning_rate in zip(progress, schedule):
prev_variational_param = variational_param
if has_log_norm:
obj_val, obj_grad, log_norm = objective_and_grad(
variational_param)
else:
obj_val, obj_grad = objective_and_grad(variational_param)
log_norm = 0
value_history.append(obj_val)
local_grad_history.append(obj_grad)
local_log_norm_history.append(log_norm)
log_norm_history.append(log_norm)
if len(local_grad_history) > window:
local_grad_history.pop(0)
local_log_norm_history.pop(0)
grad_scale = np.exp(
np.min(local_log_norm_history) -
np.array(local_log_norm_history))
scaled_grads = grad_scale[:, np.newaxis] * np.array(
local_grad_history)
accum_sum = np.sum(scaled_grads**2, axis=0)
variational_param = variational_param - curr_learning_rate * obj_grad / np.sqrt(
epsilon + accum_sum)
if i >= 3 * n_iters // 4:
variational_param_history.append(variational_param.copy())
if i % 10 == 0:
avg_loss = np.mean(value_history[max(0, i - 1000):i + 1])
progress.set_description(
'Average Loss = {:,.5g}'.format(avg_loss))
except (KeyboardInterrupt, StopIteration) as e: # pragma: no cover
# do not print log on the same line
progress.close()
finally:
progress.close()
variational_param_history = np.array(variational_param_history)
smoothed_opt_param = np.mean(variational_param_history, axis=0)
return (smoothed_opt_param, variational_param_history,
np.array(value_history), np.array(log_norm_history))
def __init__(self, mle_par, prior_par, x_mat, y_vec, y_g_vec):
self.mle_par = copy.deepcopy(mle_par)
self.prior_par = copy.deepcopy(prior_par)
self.x_mat = np.array(x_mat)
self.y_vec = np.array(y_vec)
self.y_g_vec = np.array(y_g_vec)
assert np.min(y_g_vec) == 0
assert np.max(y_g_vec) == self.mle_par['u'].size() - 1
def compute_khat_iterates(iterate_chains,
warmup=0.85,
param_idx=0,
increasing=True):
"""
Compute the khat over iterates for a variational parameter after removing warmup.
Parameters
----------
iterate_chains : multi-dimensional array, shape=(n_chains, n_iters, n_var_params)
warmup : warmup iterates
param_idx : index of the variational parameter
increasing : boolean sort array in increasing order: TRUE or decreasing order:FALSE
fraction: the fraction of iterates
Returns
-------
maximum of khat over all chains for the variational parameter param_idx
"""
chains = iterate_chains[:, :, param_idx]
n_iters = chains.shape[1]
n_chains = chains.shape[0]
k_hat_values = np.zeros(n_chains)
for i in range(n_chains):
if increasing:
sorted_chain = np.sort(chains[i, :])
else:
sorted_chain = np.sort(-chains[i, :])
ind_last = int(n_iters * warmup)
filtered_chain = sorted_chain[ind_last:]
if increasing:
filtered_chain = filtered_chain - np.min(filtered_chain)
else:
filtered_chain = filtered_chain - np.min(filtered_chain)
k_post, _ = gpdfit(filtered_chain)
k_hat_values[i] = k_post
return np.nanmax(k_hat_values)
def rand_psd(n, minew=0.1, maxew=1.):
X = np.random.randn(n,n)
S = np.dot(T_(X), X)
S = sym(S)
ew, ev = np.linalg.eigh(S)
ew -= np.min(ew)
ew /= np.max(ew)
ew *= (maxew - minew)
ew += minew
return dot3(ev, np.diag(ew), T_(ev))
def _space_constraint(self, x_in, min_dist):
x = np.nan_to_num(x_in[0:self.nturbs])
y = np.nan_to_num(x_in[self.nturbs:])
dist = [np.sqrt((x[i]-x[j])**2 + (y[i]-y[j])**2) \
for i in range(self.nturbs) \
for j in range(self.nturbs) if i != j]
return np.min(dist) - self._norm(min_dist, self.bndx_min,
self.bndx_max)
def scatter_3d_points(self, x, ax):
# set plotting limits
xmax = copy.deepcopy(np.max(x[0, :]))
xmin = copy.deepcopy(np.min(x[0, :]))
xgap = (xmax - xmin) * 0.2
xmin -= xgap
xmax += xgap
xmax1 = copy.deepcopy(np.max(x[1, :]))
xmin1 = copy.deepcopy(np.min(x[1, :]))
xgap1 = (xmax1 - xmin1) * 0.2
xmin1 -= xgap1
xmax1 += xgap1
ymax = copy.deepcopy(np.max(self.y))
ymin = copy.deepcopy(np.min(self.y))
ygap = (ymax - ymin) * 0.2
ymin -= ygap
ymax += ygap
# plot data
ax.scatter(x[0, :].flatten(),
x[1, :].flatten(),
self.y.flatten(),
color='k',
edgecolor='w',
linewidth=0.9,
s=40)
# clean up panel
ax.xaxis.pane.fill = False
ax.yaxis.pane.fill = False
ax.zaxis.pane.fill = False
ax.xaxis.pane.set_edgecolor('white')
ax.yaxis.pane.set_edgecolor('white')
ax.zaxis.pane.set_edgecolor('white')
ax.xaxis._axinfo["grid"]['color'] = (1, 1, 1, 0)
ax.yaxis._axinfo["grid"]['color'] = (1, 1, 1, 0)
ax.zaxis._axinfo["grid"]['color'] = (1, 1, 1, 0)
return xmin, xmax, xmin1, xmax1, ymin, ymax
def plot_data_and_pred(x, y, model, draw_verticals=True):
x_range = np.linspace(np.min(x), np.max(x), 100)
yhat_range = model.predict(x_range)
fig = plt.figure()
ax = fig.add_subplot(111)
ax.plot(x, y, 'o', label='observed')
ax.plot(x_range, yhat_range, 'r-', label='predicted')
if draw_verticals: # from observed value to predicted true
yhat_sparse = model.predict(x)
for x0, y0, yhat0 in zip(x, y, yhat_sparse):
ax.plot([x0, x0],[y0, yhat0],'k-')
plt.legend() #[line_pred, line_true], ['predicted', 'true'])
def plot_images(images, ax, ims_per_row=5, padding=5, digit_dimensions=(28, 28),
cmap=matplotlib.cm.binary, vmin=None, vmax=None):
"""Images should be a (N_images x pixels) matrix."""
N_images = images.shape[0]
N_rows = np.ceil(float(N_images) / ims_per_row)
pad_value = np.min(images.ravel())
concat_images = np.full(((digit_dimensions[0] + padding) * N_rows + padding,
(digit_dimensions[1] + padding) * ims_per_row + padding), pad_value)
for i in range(N_images):
cur_image = np.reshape(images[i, :], digit_dimensions)
row_ix = i // ims_per_row
col_ix = i % ims_per_row
row_start = padding + (padding + digit_dimensions[0]) * row_ix
col_start = padding + (padding + digit_dimensions[1]) * col_ix
concat_images[row_start: row_start + digit_dimensions[0],
col_start: col_start + digit_dimensions[1]] = cur_image
cax = ax.matshow(concat_images, cmap=cmap, vmin=vmin, vmax=vmax)
plt.xticks(np.array([]))
plt.yticks(np.array([]))
return cax
def plot_runtime(ex, fname, func_xvalues, xlabel, func_title=None):
results = glo.ex_load_result(ex, fname)
value_accessor = lambda job_results: job_results['time_secs']
vf_pval = np.vectorize(value_accessor)
# results['test_results'] is a dictionary:
# {'test_result': (dict from running perform_test(te) '...':..., }
times = vf_pval(results['test_results'])
repeats, _, n_methods = results['test_results'].shape
time_avg = np.mean(times, axis=0)
time_std = np.std(times, axis=0)
xvalues = func_xvalues(results)
#ns = np.array(results[xkey])
#te_proportion = 1.0 - results['tr_proportion']
#test_sizes = ns*te_proportion
line_styles = exglo.func_plot_fmt_map()
method_labels = exglo.get_func2label_map()
func_names = [f.__name__ for f in results['method_job_funcs'] ]
for i in range(n_methods):
te_proportion = 1.0 - results['tr_proportion']
fmt = line_styles[func_names[i]]
#plt.errorbar(ns*te_proportion, mean_rejs[:, i], std_pvals[:, i])
method_label = method_labels[func_names[i]]
plt.errorbar(xvalues, time_avg[:, i], yerr=time_std[:,i], fmt=fmt,
label=method_label)
ylabel = 'Time (s)'
plt.ylabel(ylabel)
plt.xlabel(xlabel)
plt.gca().set_yscale('log')
plt.xlim([np.min(xvalues), np.max(xvalues)])
plt.xticks( xvalues, xvalues)
plt.legend(loc='best')
title = '%s. %d trials. '%( results['prob_label'],
repeats ) if func_title is None else func_title(results)
plt.title(title)
#plt.grid()
return results
def fun(x): return to_scalar(np.min(x)) d_fun = lambda x : to_scalar(grad(fun)(x))
def fun(x): return to_scalar(np.min(x, axis=1, keepdims=True)) d_fun = lambda x : to_scalar(grad(fun)(x))
def polyinterp(points, doPlot=None, xminBound=None, xmaxBound=None):
""" polynomial interpolation
Parameters
----------
points: shape(pointNum, 3), three columns represents x, f, g
doPolot: set to 1 to plot, default 0
xmin: min value that brackets minimum (default: min of points)
xmax: max value that brackets maximum (default: max of points)
set f or g to sqrt(-1)=1j if they are not known
the order of the polynomial is the number of known f and g values minus 1
Returns
-------
minPos:
fmin:
"""
if doPlot == None:
doPlot = 0
nPoints = points.shape[0]
order = np.sum(np.imag(points[:, 1:3]) == 0) -1
# code for most common case: cubic interpolation of 2 points
if nPoints == 2 and order == 3 and doPlot == 0:
[minVal, minPos] = [np.min(points[:,0]), np.argmin(points[:,0])]
notMinPos = 1 - minPos
d1 = points[minPos,2] + points[notMinPos,2] - 3*(points[minPos,1]-\
points[notMinPos,1])/(points[minPos,0]-points[notMinPos,0])
t_d2 = d1**2 - points[minPos,2]*points[notMinPos,2]
if t_d2 > 0:
d2 = np.sqrt(t_d2)
else:
d2 = np.sqrt(-t_d2) * np.complex(0,1)
if np.isreal(d2):
t = points[notMinPos,0] - (points[notMinPos,0]-points[minPos,0])*\
((points[notMinPos,2]+d2-d1)/(points[notMinPos,2]-\
points[minPos,2]+2*d2))
minPos = np.min([np.max([t,points[minPos,0]]), points[notMinPos,0]])
else:
minPos = np.mean(points[:,0])
fmin = minVal
return (minPos, fmin)
xmin = np.min(points[:,0])
xmax = np.max(points[:,0])
# compute bounds of interpolation area
if xminBound == None:
xminBound = xmin
if xmaxBound == None:
xmaxBound = xmax
# constraints based on available function values
A = np.zeros((0, order+1))
b = np.zeros((0, 1))
for i in range(nPoints):
if np.imag(points[i,1]) == 0:
constraint = np.zeros(order+1)
for j in np.arange(order,-1,-1):
constraint[order-j] = points[i,0]**j
A = np.vstack((A, constraint))
b = np.append(b, points[i,1])
# constraints based on availabe derivatives
for i in range(nPoints):
if np.isreal(points[i,2]):
constraint = np.zeros(order+1)
for j in range(1,order+1):
constraint[j-1] = (order-j+1)* points[i,0]**(order-j)
A = np.vstack((A, constraint))
b = np.append(b,points[i,2])
# find interpolating polynomial
params = np.linalg.solve(A, b)
# compute critical points
dParams = np.zeros(order)
for i in range(params.size-1):
dParams[i] = params[i] * (order-i)
if np.any(np.isinf(dParams)):
cp = np.concatenate((np.array([xminBound, xmaxBound]), points[:,0]))
else:
cp = np.concatenate((np.array([xminBound, xmaxBound]), points[:,0], \
np.roots(dParams)))
# test critical points
fmin = np.infty;
minPos = (xminBound + xmaxBound)/2.
for xCP in cp:
if np.imag(xCP) == 0 and xCP >= xminBound and xCP <= xmaxBound:
fCP = np.polyval(params, xCP)
if np.imag(fCP) == 0 and fCP < fmin:
minPos = np.double(np.real(xCP))
fmin = np.double(np.real(fCP))
# plot situation (omit this part for now since we are not going to use it
# anyway)
return (minPos, fmin)
def fun(x): return to_scalar(np.min(np.array([[x, x, x],
[x, 0.5, 0.5],
[0.5, 0.5, 0.5],
[x, x, 0.5]]), axis=0))
d_fun = lambda x : to_scalar(grad(fun)(x))
def _read_kurucz_spec(f):
"""
Read Kurucz spectra that have been precomputed
Args:
f (string) : path to the file to be read
Returns:
new_vel (real array) : velocity axis in km/s
spectrum (real array) : spectrum for each velocity bin
"""
f = open(f, "rb")
res = f.read()
n_chunk = struct.unpack('i',res[0:4])
freq = []
stokes = []
cont = []
left = 4
for i in range(n_chunk[0]):
right = left + 4
n = struct.unpack('i',res[left:right])
left = right
right = left + 4
nmus = struct.unpack('i',res[left:right])
left = right
right = left + 8*n[0]
t1 = np.asarray(struct.unpack('d'*n[0],res[left:right]))
freq.append(t1)
left = right
right = left + 8*n[0]*nmus[0]
t2 = np.asarray(struct.unpack('d'*n[0]*nmus[0],res[left:right])).reshape((n[0],nmus[0]))
stokes.append(t2)
left = right
right = left + 8*n[0]*nmus[0]
t2 = np.asarray(struct.unpack('d'*n[0]*nmus[0],res[left:right])).reshape((n[0],nmus[0]))
cont.append(t2)
left = right
freq = np.concatenate(freq)
stokes = np.concatenate(stokes)
cont = np.concatenate(cont)
ind = np.argsort(freq)
freq = freq[ind]
stokes = stokes[ind]
cont = cont[ind]
wavelength = const.c.to('cm/s').value / freq
mean_wavelength = np.mean(wavelength)
vel = (wavelength - mean_wavelength) / mean_wavelength * const.c.to('km/s').value
nl, nmus = stokes.shape
# Reinterpolate in a equidistant velocity axis
new_vel = np.linspace(np.min(vel), np.max(vel), nl)
for i in range(nmus):
interpolator = scipy.interpolate.interp1d(vel, stokes[:,i], kind='linear')
stokes[:,i] = interpolator(new_vel)
return new_vel, wavelength, stokes
# using sklearn #
################################################################
N = 500
#features,labels = ds.make_classification(n_samples = N,n_features = 2,n_informative = 2,n_redundant = 0,n_clusters_per_class = 1,class_sep = 2,shift = 2.2)
features,labels = ds.make_circles(n_samples = N)
#features,labels = ds.make_moons(n_samples = N)
labels[labels == 0] = -1
features = auto_np.array(features) * 4.0
labels = auto_np.array(labels).reshape(features.shape[0],1)
return features,labels
if __name__ == '__main__' :
features,labels = gen_test_data()
optimized_params = neural_net_train(features,labels,num_iter = 2000)
#optimized_params = neural_net_train(features,labels,num_iter = 1000,opt_method = 'stepest')
min_x = auto_np.min(features[:,0])
max_x = auto_np.max(features[:,0])
min_y = auto_np.min(features[:,1])
max_y = auto_np.max(features[:,1])
xx,yy = np.meshgrid(np.linspace(min_x,max_x,200),np.linspace(min_y,max_y,200))
predict_features = np.c_[xx.ravel(),yy.ravel()]
predict_labels = neural_net_predict(optimized_params,predict_features).reshape(xx.shape)
cs = plt.contour(xx,yy,predict_labels)
plt.clabel(cs,inline = 1,fontsize = 10)
plt.scatter(features[labels[:,0] == 1,0],features[labels[:,0] == 1,1],c = 'red')
plt.scatter(features[labels[:,0] == -1,0],features[labels[:,0] == -1,1],c = 'cyan')
plt.show()
# phi - [0, 180], transformation log (phi / (180 - phi))
#
######################################################################
import CelestePy.util.data as du
from sklearn.linear_model import LinearRegression
coadd_df = du.load_celeste_dataframe("../../data/stripe_82_dataset/coadd_catalog_from_casjobs.fit")
# make star => radial extent proposal
star_res = coadd_df.gal_arcsec_scale[ coadd_df.is_star ].values
star_res = np.clip(star_res, 1e-8, np.inf)
star_res_proposal = fit_mog(np.log(star_res).reshape((-1,1)), max_comps = 20, mog_class = MixtureOfGaussians)
with open('star_res_proposal.pkl', 'wb') as f:
pickle.dump(star_res_proposal, f)
if False:
xgrid = np.linspace(np.min(np.log(star_res)), np.max(np.log(star_res)), 100)
lpdf = star_res_proposal.logpdf(xgrid.reshape((-1,1)))
plt.plot(xgrid, np.exp(lpdf))
plt.hist(np.log(star_res), 25, normed=True)
plt.hist(np.log(star_res), 25, normed=True, alpha=.24)
plt.hist(star_res_proposal.rvs(684).flatten(), 25, normed=True, alpha=.24)
# make star fluxes => gal fluxes for tars
colors = ['ug', 'gr', 'ri', 'iz']
star_mags = np.array([du.colors_to_mags(r, c)
for r, c in zip(coadd_df.star_mag_r.values,
coadd_df[['star_color_%s'%c for c in colors]].values)])
gal_mags = np.array([du.colors_to_mags(r, c)
for r, c in zip(coadd_df.gal_mag_r.values,
coadd_df[['gal_color_%s'%c for c in colors]].values)])
