def sample_mean_cov_from_deep_gp(all_params, X, with_noise = False):
predict = predict_funcs_with_noise if with_noise else predict_layer_funcs
X_star = X
layer_params, x0, y0 = unpack_all_params(all_params)
n_layers = len(x0)
for layer in xrange(n_layers):
layer_mean, layer_cov = predict[layer](layer_params[layer],np.atleast_2d(x0[layer]).T, y0[layer],X_star)
X_star = np.atleast_2d(sample_from_mvn(layer_mean, layer_cov)).T
return layer_mean,layer_cov
def plot_isocontours(ax, func, xlimits=[-2, 2], ylimits=[-4, 2], numticks=101):
x = np.linspace(*xlimits, num=numticks)
y = np.linspace(*ylimits, num=numticks)
X, Y = np.meshgrid(x, y)
zs = func(np.concatenate([np.atleast_2d(X.ravel()), np.atleast_2d(Y.ravel())]).T)
Z = zs.reshape(X.shape)
plt.contour(X, Y, Z)
ax.set_yticks([])
ax.set_xticks([])
def evaluate_prior(all_params): # clean up code so we don't compute matrices twice
layer_params, x0, y0 = unpack_all_params(all_params)
log_prior = 0
for layer in xrange(n_layers):
#import pdb; pdb.set_trace()
mean, cov_params, noise_scale = unpack_kernel_params(layer_params[layer])
cov_y_y = covariance_function(cov_params, np.atleast_2d(x0[layer]).T, np.atleast_2d(x0[layer]).T) + noise_scale * np.eye(len(y0[layer]))
log_prior += mvn.logpdf(y0[layer],np.ones(len(cov_y_y))*mean,cov_y_y+np.eye(len(cov_y_y))*1e-6*np.max(np.diag(cov_y_y)))
return log_prior
def mvt_logpdf(x, mu, Li, df):
dim = Li.shape[0]
Ki = np.dot(Li.T, Li)
#determinant is just multiplication of diagonal elements of cholesky
logdet = 2*log(1./np.diag(Li)).sum()
lpdf_const = (gammaln((df + dim) / 2)
-(gammaln(df/2)
+ (log(df)+log(np.pi)) * dim*0.5
+ logdet * 0.5)
)
x = np.atleast_2d(x)
if x.shape[1] != mu.size:
x = x.T
assert(x.shape[1] == mu.size
or x.shape[0] == mu.size)
d = (x - mu.reshape((1 ,mu.size))).T
Ki_d_scal = np.dot(Ki, d) /df #vector
d_Ki_d_scal_1 = diag_dot(d.T, Ki_d_scal) + 1. #scalar
res_pdf = (lpdf_const
- 0.5 * (df + dim) * np.log(d_Ki_d_scal_1)).flatten()
if res_pdf.size == 1:
res_pdf = np.float(res_pdf)
return res_pdf
def fit(cls, samples, return_instance = False): # observations expected in rows
mu = samples.mean(0)
var = np.atleast_2d(np.cov(samples, rowvar = 0))
if return_instance:
return mvnorm(mu, var)
else:
return (mu, var)
def log_pdf_and_grad(self, x, pdf = True, grad = True, T = np):
assert(pdf or grad)
if T == np:
x = np.atleast_2d(x)
if x.shape[1] != self.mu.size:
x = x.T
assert(np.sum(np.array(x.shape) == self.mu.size)>=1)
d = (x - self.mu.reshape((1 ,self.mu.size))).T
Ki_d = T.dot(self.Ki, d) #vector
if pdf:
# vector times vector
res_pdf = (self.lpdf_const - 0.5 * diag_dot(d.T, Ki_d)).T
if res_pdf.size == 1:
res_pdf = res_pdf.reshape(res_pdf.size)[0]
if not grad:
return res_pdf
if grad:
# nothing
res_grad = - Ki_d.T #.flat[:]
if res_grad.shape[0] <= 1:
res_grad = res_grad.flatten()
if not pdf:
return res_grad
return (res_pdf, res_grad)
def grad_func(*args):
inp = anp.array(anp.broadcast_arrays(*args))
result = anp.atleast_2d(elementwise_grad(argwrapper)(inp))
# Put 'gradient' axis at end
axes = list(range(len(result.shape)))
result = result.transpose(*chain(axes[1:], [axes[0]]))
return result
def __init__(self,
state_dim: np.ndarray,
target_state: np.ndarray = None,
weights: np.ndarray = None,
cost_width: np.ndarray = None):
"""
Initialize saturated loss function
:param state_dim: state dimensionality
:param target_state: target state which should be reached
:param weights: weight matrix
:param cost_width: TODO what is this
"""
self.state_dim = state_dim
# set target state to all zeros if not other specified
self.target_state = np.atleast_2d(
np.zeros(self.state_dim) if target_state is None else target_state)
# weight matrix
self.weights = np.identity(
self.state_dim) if weights is None else weights
# -----------------------------------------------------
# This is only useful if we have any penalties etc.
self.cost_width = np.array([1]) if cost_width is None else cost_width
def __call__(self, log_hyperparams, x, z=None):
log_hyperparams = np.atleast_2d(log_hyperparams)
left, right = self.sub
split = left.n_hyperparams(x)
return left(log_hyperparams[:, :split], x, z) + right(
log_hyperparams[:, split:], x, z)
def ppf(self, component_cum_prob, eig=True):
assert(component_cum_prob.shape[1] == self.get_num_unif())
rval = []
for i in range(component_cum_prob.shape[0]):
r = component_cum_prob[i,:]
comp = self.dist_cat.ppf(r[0])
rval.append(self.comp_dist[comp].ppf(np.atleast_2d(r[1:]), eig=True))
return np.array(rval).reshape((component_cum_prob.shape[0], self.dim))
def equa2pixel(self, s_equa):
if self.use_wcs:
print "using wcs in equa2pixel"
return self.wcs.wcs_world2pix(np.atleast_2d(s_equa), 0).squeeze()
phi1rad = self.phi_n[1] / 180. * np.pi
s_iwc = np.array([ (s_equa[0] - self.phi_n[0]) * np.cos(phi1rad),
(s_equa[1] - self.phi_n[1]) ])
s_pix = np.dot(self.Ups_n_inv, s_iwc) + self.rho_n
return s_pix
def callback(params):
print("Log likelihood {}, Squared Error {}".format(-objective(params),squared_error(params)))
layer_params, x0, y0 = unpack_all_params(params)
# Show posterior marginals.
plot_xs = np.reshape(np.linspace(-5, 5, 300), (300,1))
plot_full_gp(ax_end_to_end, params, plot_xs)
if n_layers == 1:
ax_end_to_end.plot(x0[0],y0[0], 'ro')
else:
hidden_mean, hidden_cov = predict_layer_funcs[0](layer_params[0], np.atleast_2d(x0[0]).T, y0[0], plot_xs)
plot_gp(ax_x_to_h, x0[0], y0[0], hidden_mean, hidden_cov, plot_xs)
ax_x_to_h.set_title("X to hiddens, with inducing points")
y_mean, y_cov = predict_layer_funcs[1](layer_params[1], np.atleast_2d(x0[1]).T, y0[1], plot_xs)
plot_gp(ax_h_to_y, x0[1], y0[1], y_mean, y_cov, plot_xs)
ax_h_to_y.set_title("hiddens to layer 2, with inducing points")
plt.draw()
plt.pause(1.0/60.0)
def mvt_ppf(component_cum_prob, mu, L, df):
from scipy.stats import norm, chi2
mu = np.atleast_1d(mu).flatten()
assert(component_cum_prob.shape[1] == mu.size+1)
L = np.atleast_2d(L)
rval = []
for r in range(component_cum_prob.shape[0]):
samp_mvn_0mu = L.dot(norm.ppf(component_cum_prob[r, :-1]))
samp_chi2 = chi2.ppf(component_cum_prob[r, -1], df)
samp_mvt_0mu = samp_mvn_0mu * np.sqrt(df / samp_chi2)
rval.append(mu + samp_mvt_0mu)
return np.array(rval)
def callback(params):
print("Log marginal likelihood {}".format(log_marginal_likelihood(params)))
# Show posterior marginals.
plot_xs = np.reshape(np.linspace(-5, 5, 300), (300,1))
pred_mean, pred_cov = combined_predict_fun(params, X, y, plot_xs)
plot_gp(ax_end_to_end, X, y, pred_mean, pred_cov, plot_xs)
ax_end_to_end.set_title("X to y")
layer1_params, layer2_params, hiddens = unpack_all_params(params)
h_star_mean, h_star_cov = predict_layer_funcs[0](layer1_params, X, hiddens, plot_xs)
y_star_mean, y_star_cov = predict_layer_funcs[0](layer2_params, np.atleast_2d(hiddens).T, y, plot_xs)
plot_gp(ax_x_to_h, X, hiddens, h_star_mean, h_star_cov, plot_xs)
ax_x_to_h.set_title("X to hiddens")
plot_gp(ax_h_to_y, np.atleast_2d(hiddens).T, y, y_star_mean, y_star_cov, plot_xs)
ax_h_to_y.set_title("hiddens to y")
plt.draw()
plt.pause(1.0/60.0)
def choose_next_point(domain_min, domain_max, acquisition_function, num_tries=15, rs=npr.RandomState(0)):
"""Uses gradient-based optimization to find next query point."""
init_points = rs.rand(num_tries, D) * (domain_max - domain_min) + domain_min
grad_obj = value_and_grad(lambda x: -acquisition_function(x))
def optimize_point(init_point):
print('.', end='')
result = minimize(grad_obj, x0=init_point, jac=True, method='L-BFGS-B',
options={'maxiter': 10}, bounds=list(zip(domain_min, domain_max)))
return result.x, acquisition_function(result.x)
optimzed_points, optimized_values = list(zip(*list(map(optimize_point, init_points))))
print()
best_ix = np.argmax(optimized_values)
return np.atleast_2d(optimzed_points[best_ix])
def mog_like(x, means, icovs, dets, pis):
""" compute the log likelihood according to a mixture of gaussians
with means = [mu0, mu1, ... muk]
icovs = [C0^-1, ..., CK^-1]
dets = [|C0|, ..., |CK|]
pis = [pi1, ..., piK] (sum to 1)
at locations given by x = [x1, ..., xN]
"""
xx = np.atleast_2d(x)
centered = xx[:,:,np.newaxis] - means.T[np.newaxis,:,:]
solved = np.einsum('ijk,lji->lki', icovs, centered)
logprobs = -0.5*np.sum(solved * centered, axis=1) - np.log(2*np.pi) - 0.5*np.log(dets) + np.log(pis)
logprob = scpm.logsumexp(logprobs, axis=1)
if len(x.shape) == 1:
return np.exp(logprob[0])
else:
return np.exp(logprob)
def mog_logmarglike(x, means, covs, pis, ind=0):
""" marginal x or y (depending on ind) """
K = pis.shape[0]
xx = np.atleast_2d(x)
centered = xx.T - means[:,ind,np.newaxis].T
logprobs = []
for kk in xrange(K):
quadterm = centered[:,kk] * centered[:,kk] * (1./covs[kk,ind,ind])
logprobsk = -.5*quadterm - .5*np.log(2*np.pi) \
-.5*np.log(covs[kk,ind,ind]) + np.log(pis[kk])
logprobs.append(np.squeeze(logprobsk))
logprobs = np.array(logprobs)
logprob = scpm.logsumexp(logprobs, axis=0)
if np.isscalar(x):
return logprob[0]
else:
return logprob
def __init__(self, mu, K, Ki = None, logdet_K = None, L = None):
mu = np.atleast_1d(mu).flatten()
K = np.atleast_2d(K)
assert(np.prod(mu.shape) == K.shape[0] )
assert(K.shape[0] == K.shape[1])
self.mu = mu
self.K = K
(val, vec) = np.linalg.eigh(K)
idx = np.arange(mu.size-1,-1,-1)
(self.eigval, self.eigvec) = (np.diag(val[idx]), vec[:,idx])
self.eig = self.eigvec.dot(np.sqrt(self.eigval))
self.dim = K.shape[0]
#(self.Ki, self.logdet) = (np.linalg.inv(K), np.linalg.slogdet(K)[1])
(self.Ki, self.L, self.Li, self.logdet) = pdinv(K)
self.lpdf_const = -0.5 *np.float(self.dim * np.log(2 * np.pi)
+ self.logdet)
def __init__(self, mu, K, df, Ki = None, logdet_K = None, L = None):
mu = np.atleast_1d(mu).flatten()
K = np.atleast_2d(K)
assert(np.prod(mu.shape) == K.shape[0] )
assert(K.shape[0] == K.shape[1])
self.mu = mu
self.K = K
self.df = df
self._freeze_chi2 = stats.chi2(df)
self.dim = K.shape[0]
self._df_dim = self.df + self.dim
#(self.Ki, self.logdet) = (np.linalg.inv(K), np.linalg.slogdet(K)[1])
(self.Ki, self.L, self.Li, self.logdet) = pdinv(K)
self.lpdf_const = np.float(gammaln((self.df + self.dim) / 2)
-(gammaln(self.df/2)
+ (log(self.df)+log(np.pi)) * self.dim*0.5
+ self.logdet * 0.5)
)
def gmm_logprob(x, ws, mus, sigs, invsigs=None, logdets=None):
""" Gaussian Mixture Model likelihood
Input:
- x = N x D array of data (N iid)
- ws = K length vector that sums to 1, mixing weights
- mus = K x D array of mixture component means
- sigs = K x D x D array of mixture component covariances
- invsigs = K x D x D array of mixture component covariance inverses
- logdets = K array of mixture component covariance logdets
Output:
- N length array of log likelihood values
TODO: speed this up
"""
if sigs is None:
assert invsigs is not None and logdets is not None, \
"need sigs if you don't include logdets and invsigs"
# compute invsigs if needed
if invsigs is None:
invsigs = np.array([np.linalg.inv(sig) for sig in sigs])
logdets = np.array([np.linalg.slogdet(sig)[1] for sig in sigs])
# compute each gauss component separately
xx = np.atleast_2d(x)
centered = xx[:,:,np.newaxis] - mus.T[np.newaxis,:,:]
solved = np.einsum('ijk,lji->lki', invsigs, centered)
logprobs = -0.5*np.sum(solved * centered, axis=1) - \
np.log(2*np.pi) - 0.5*logdets + np.log(ws)
logprob = scpm.logsumexp(logprobs, axis=1)
if len(x.shape) == 1:
return logprob[0]
else:
return logprob
def plot_deep_gp(ax, params, plot_xs):
ax.cla()
rs = npr.RandomState(0)
sampled_means_and_covs = [sample_mean_cov_from_deep_gp(params, plot_xs, rs = rs, with_noise = False, FITC = False) for i in xrange(n_samples_to_plot)]
sampled_means, sampled_covs = zip(*sampled_means_and_covs)
avg_pred_mean = np.mean(sampled_means, axis = 0)
avg_pred_cov = np.mean(sampled_covs, axis = 0)
avg_pred_cov = avg_pred_cov+np.sum(np.array([np.dot(np.atleast_2d(sampled_means[i]-avg_pred_mean).T,np.atleast_2d((sampled_means[i]-avg_pred_mean))) for i in xrange(n_samples_to_plot)]),axis = 0)/n_samples
marg_std = np.sqrt(np.diag(avg_pred_cov))
if n_samples_to_plot > 19:
ax.fill(np.concatenate([plot_xs, plot_xs[::-1]]),
np.concatenate([avg_pred_mean - 1.96 * marg_std,
(avg_pred_mean + 1.96 * marg_std)[::-1]]),
alpha=.15, fc='Blue', ec='None')
ax.plot(plot_xs, avg_pred_mean, 'b')
sampled_funcs = np.array([rs.multivariate_normal(mean, cov*(random)) for mean,cov in sampled_means_and_covs])
ax.plot(plot_xs,sampled_funcs.T)
ax.plot(X, y, 'kx')
#ax.set_ylim([-1.5,1.5])
ax.set_xticks([])
ax.set_yticks([])
ax.set_title("Full Deep GP, inputs to outputs")
def sample_from_mvn(mu, sigma): # make sure we return 2d, also make sure data is 2d
rs = npr.RandomState(0)
return np.atleast_2d(np.dot(np.linalg.cholesky(sigma+1e-6*np.eye(len(sigma))*np.max(np.diag(sigma))),rs.randn(len(sigma)))+mu if random == 1 else mu).T
def combined_predict_fun(all_params, X, y, xs):
layer1_params, layer2_params, hiddens = unpack_all_params(all_params)
h_star_mean, h_star_cov = predict_layer1(layer1_params, X, hiddens, xs)
y_star_mean, y_star_cov = predict_layer2(layer2_params, np.atleast_2d(hiddens).T, y, np.atleast_2d(h_star_mean).T)
return y_star_mean, y_star_cov
def log_marginal_likelihood(all_params):
layer1_params, layer2_params, h = unpack_all_params(all_params)
return log_marginal_likelihood_layer1(layer1_params, X, h) + \
log_marginal_likelihood_layer2(layer2_params, np.atleast_2d(h).T, y)
def acquisition_function(xstar):
xstar = np.atleast_2d(xstar) # To work around a bug in scipy.minimize
mean, std = predict_func(xstar)
return expected_new_max(mean, std, defaultmax(y))
def fit(cls, samples, return_instance = False): # observations expected in rows
raise(NotImplementedError())
mu = samples.mean(0)
return (mu, np.atleast_2d(np.cov(samples, rowvar = 0)))