def get_general_node_params(x, lds):
T, p = x.shape
C, sigma_obs = lds[-2:]
J, Jzx, Jxx, logZ = pair_mean_to_natural(C, sigma_obs)
h = np.einsum('tzx,tx->tz', Jzx, x)
logZ += np.einsum('ti,tij,tj->t', x, Jxx, x) - p/2.*np.log(2*np.pi)
return J, h, logZ
def log_marginal_likelihood(self, theta):
x_train = self.X_train_
y_train = self.y_train_
if np.ndim == 1:
y_train = y_train[:, np.newaxis]
# Gather hyper parameters
signal_variance, noise_likelihood, length_scale = \
self._get_kernel_params(theta)
signal_variance = np.exp(signal_variance)
noise_likelihood = np.exp(noise_likelihood)
length_scale = np.exp(length_scale)
n_samples = x_train.shape[0]
# train kernel
K = self.rbf_covariance(x_train,
length_scale=length_scale,
signal_variance=signal_variance)
K += noise_likelihood * np.eye(n_samples)
L = np.linalg.cholesky(K + self.jitter * np.eye(n_samples))
weights = np.linalg.solve(L.T, np.linalg.solve(L, y_train))
log_likelihood_dims = -0.5 * np.einsum("ik,ik->k", y_train, weights)
log_likelihood_dims -= np.log(np.diag(L)).sum()
log_likelihood_dims -= (K.shape[0] / 2) * np.log(2 * np.pi)
log_likelihood = log_likelihood_dims.sum(-1)
return -log_likelihood
def generalized_outer_product(mat):
if len(mat.shape) == 1:
return np.outer(mat, mat)
elif len(mat.shape) == 2:
return np.einsum('ij,ik->ijk', mat, mat)
else:
raise ArithmeticError
def predict(self,
user_id_N,
item_id_N,
mu=None,
b_per_user=None,
c_per_item=None,
U=None,
V=None):
''' Predict ratings at specific user_id, item_id pairs
Args
----
user_id_N : 1D array, size n_examples
Specific user_id values to use to make predictions
item_id_N : 1D array, size n_examples
Specific item_id values to use to make predictions
Each entry is paired with the corresponding entry of user_id_N
Returns
-------
yhat_N : 1D array, size n_examples
Scalar predicted ratings, one per provided example.
Entry n is for the n-th pair of user_id, item_id values provided.
'''
# TODO: Update with actual prediction logic
N = user_id_N.size
yhat_N = ag_np.ones(N)*mu+b_per_user[user_id_N]+c_per_item[item_id_N]\
+ag_np.einsum('ij, ij->i', U[user_id_N], V[item_id_N])
return yhat_N
def _ll(self, m, p, a, xn, xln, **kwargs):
"""Computation of log likelihood
Dimensions
----------
m : n_unique x n_features
p : n_unique x n_features x n_features
a : n_unique x n_lags (shared_alpha=F)
OR 1 x n_lags (shared_alpha=T)
xn: N x n_features
xln: N x n_features x n_lags
"""
samples = xn.shape[0]
xn = xn.reshape(samples, 1, self.n_features)
m = m.reshape(1, self.n_unique, self.n_features)
det = np.linalg.det(np.linalg.inv(p))
det = det.reshape(1, self.n_unique)
lagged = np.dot(xln, a.T) # NFU
lagged = np.swapaxes(lagged, 1, 2) # NUF
xm = xn-(lagged + m)
tem = np.einsum('NUF,UFX,NUX->NU', xm, p, xm)
res = (-self.n_features/2.0)*np.log(2*np.pi) - 0.5*tem - 0.5*np.log(det)
return res
def magcal_residual2(a, mb, gb, gs):
""" 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)
dt = TS[1:] - TS[:-1]
w = gs * (GYRO[1:] - gb)
C = so3.tensorexp(w.T * dt)
rot_mag = np.einsum('ijl,lj->li', C, mag[:-1])
return np.mean(np.abs(1 - np.einsum('ji,ji->j', mag[1:], rot_mag)))
def _ll(self, m, p, a, xn, xln, **kwargs):
"""Computation of log likelihood
Dimensions
----------
m : n_unique x n_features
p : n_unique x n_features x n_features
a : n_unique x n_lags (shared_alpha=F)
OR 1 x n_lags (shared_alpha=T)
xn: N x n_features
xln: N x n_features x n_lags
"""
samples = xn.shape[0]
xn = xn.reshape(samples, 1, self.n_features)
m = m.reshape(1, self.n_unique, self.n_features)
det = np.linalg.det(np.linalg.inv(p))
det = det.reshape(1, self.n_unique)
lagged = np.dot(xln, a.T) # NFU
lagged = np.swapaxes(lagged, 1, 2) # NUF
xm = xn-(lagged + m)
tem = np.einsum('NUF,UFX,NUX->NU', xm, p, xm)
# TODO division in gamma function
res = np.log(gamma((self.degree_freedom + self.n_features)/2)) - \
np.log(gamma(self.degree_freedom/2)) - (self.n_features/2.0) * \
np.log(self.degree_freedom) - \
(self.n_features/2.0) * np.log(np.pi) - 0.5 * np.log(det) - \
((self.degree_freedom + self.n_features) / 2.0) * \
np.log(1 + (1/self.degree_freedom) * tem)
return res
def predictions(weights, inputs):
inputs = np.expand_dims(inputs, 0)
for W, b in unpack_layers(weights):
outputs = np.einsum('mnd,mdo->mno', inputs, W) + b
inputs = nonlinearity(outputs)
#return outputs - logsumexp(outputs, axis=1, keepdims=True)
return outputs
def construct_sphere(self,
θ: "Model parameters" = None
) -> "Plots an unit sphere":
if θ is None:
θ = self.θ_bf
c, c0, ωv, A = self.compute_ω_sphere(θ, Npoints=10)
w, v = self.metric_eigenproblem(self.g(θ))
f, ax = plt.subplots(1, 3, figsize=(23, 7))
ax[0].plot(c[0], c[1], 'k', ls='-', zorder=0)
ax[1].plot(c0[0], c0[1], 'k', zorder=0)
im0 = ax[0].scatter(c[0], c[1], c=ωv / (2 * np.pi), zorder=1, s=10**2)
ax[0].set_title('Standardni sustav')
ax[1].set_title('Lokalni inercijalni sustav')
im1 = ax[1].scatter(c0[0],
c0[1],
c=ωv / (2 * np.pi),
zorder=1,
s=10**2)
e, ei = self.find_tetrad(self.θ_bf)
for i in range(2):
vn = v[:, i]
vn = vn / np.sqrt(np.einsum('i,ij,j', vn, self.g(θ), vn))
vninv = np.einsum('im,m', e, vn)
ax[0].plot([0, vn[0]], [0, vn[1]],
color='C%d' % i,
lw=3,
label='$\lambda=%e$\n' % w[i] +
'$\omega=%e$' % self.external_ωv(θ, vn))
ax[1].plot(
[0, vninv[0]],
[0, vninv[1]],
'C%d' % i,
lw=3,
)
ax[0].legend(loc=1)
f.colorbar(im0, ax=ax[0]).set_label('$\omega/v/(2\pi)$')
f.colorbar(im1, ax=ax[1]).set_label('$\omega/v/(2\pi)$')
ax[0].set_xlabel(r'$\dot\theta^{\mu=0}$')
ax[0].set_xlabel(r'$\dot\theta^{\mu=1}$')
ax[1].set_xlabel(r'$\dot\theta^{i=0}$')
ax[1].set_xlabel(r'$\dot\theta^{i=1}$')
ax[2].plot(A[0], ωv / (2 * np.pi))
ax[2].set_xlabel(r'$\varphi^0$')
ax[2].set_ylabel('$\omega/v/(2\pi)$')
f.tight_layout()
f.savefig(NAME + '_omegas.pdf')
return f
def compute_approximate_gradients(self, nvalid=0):
if self.layer_plastic[0]:
self.sleep_data["fsuff0"] = self.model.dists[0].suff(
self.sleep_data["x0"])
self.reg.train_weights(self.sleep_data,
"x0->fsuff0",
nvalid=nvalid)
for i in range(1, self.model.depth):
if not self.layer_plastic[i]:
continue
n = self.nsleep
self.sleep_data["dnorm%d" % i] = self.model.dists[i].dnorm(
self.sleep_data["x%d" % (i - 1)])
self.reg.train_weights(self.sleep_data,
"x%d->dnorm%d" % (i - 1, i),
nvalid=nvalid)
if i == (self.model.depth - 1):
data = self.sleep_data
lam = self.reg.lam
z_feat = self.reg.transform_data(
data, ["x%d" % (i - 1)])["x%d" % (i - 1)]
x_suff = self.model.dists[-1].suff(data["x%d" % i])
x = np.einsum("ij,ik->ijk", z_feat,
x_suff).reshape(self.nsleep, -1)
y = z_feat
f = self.model.dists[-1].dlogp(data["x%d" % (i - 1)],
data["x%d" % (i)])
Mx = np.eye(self.nsleep) - x.dot(
np.linalg.solve(
x.T.dot(x) + np.eye(x.shape[1]) * lam, x.T))
My = np.eye(self.nsleep) - y.dot(
np.linalg.solve(
y.T.dot(y) + np.eye(y.shape[1]) * lam, y.T))
A = np.linalg.solve(
(x.T.dot(My).dot(x)) + np.eye(x.shape[1]) * lam,
x.T.dot(My).dot(f))
B = -np.linalg.solve(
(y.T.dot(Mx).dot(y)) + np.eye(y.shape[1]) * lam,
y.T.dot(Mx).dot(f))
self.reg.Ws["A"] = A
self.reg.Ws["B"] = B
else:
self.sleep_data["dnatsuff%d" %
i] = self.model.dists[i].dnatsuff(
self.sleep_data["x%d" % (i - 1)],
self.sleep_data["x%d" % (i)])
self.reg.train_weights(self.sleep_data,
"x%d->dnatsuff%d" % (i, i),
nvalid=nvalid)
def smooth(self, gamma, x, u):
mean = []
for _x, _u, _gamma in zip(x, u, gamma):
_mu = np.zeros((len(_x) - 1, self.nb_states, self.dm_obs))
for k in range(self.nb_states):
_mu[:, k, :] = self.mean(k, _x[:-1, :], _u[:-1, :self.dm_act])
mean.append(np.einsum('nk,nkl->nl', _gamma[1:, ...], _mu))
return mean
def weingarten_map(self, X, roll=True):
g = np.rollaxis(self.metric_tensor(X), 2)
h = np.rollaxis(self.shape_tensor(X), 2)
res = np.einsum("aio, abi -> abo", np.linalg.inv(g), h)
if roll:
return np.rollaxis(res, 0, 3)
else:
return res
def devpays(mix):
"""Compute the dev pays"""
profs = sample_profs(mix)
payoffs = model_pays(profs)
numer = rep(anp.prod(mix**profs, 2))
denom = const_weights(profs, mix)
weights = numer / denom / learn.num_samples
return anp.einsum('ij,ij->j', weights, payoffs)
def update_hidden(self, weights, input, hidden, cells):
concated_input = agnp.concatenate((input, hidden), axis=2)
W_change, b_change = self.unpack_change_params(weights)
change = agnp.tanh(
agnp.einsum('pdh,pnd->pnh', W_change, concated_input) + b_change)
W_forget, b_forget = self.unpack_forget_params(weights)
forget = self.hidden_nonlinearity(
agnp.einsum('pdh,pnd->pnh', W_forget, concated_input) + b_forget)
W_ingate, b_ingate = self.unpack_ingate_params(weights)
ingate = self.hidden_nonlinearity(
agnp.einsum('pdh,pnd->pnh', W_ingate, concated_input) + b_ingate)
W_outgate, b_outgate = self.unpack_outgate_params(weights)
outgate = self.hidden_nonlinearity(
agnp.einsum('pdh,pnd->pnh', W_outgate, concated_input) + b_outgate)
cells = cells * forget + ingate * change
hidden = outgate * agnp.tanh(cells)
return hidden, cells
def log_emission_probs(tau, mu_eta, mu_c, Sig_c, xi):
"""
Calculate log psi, where psi \propto p(obs|z)
"""
xi1 = xi[:, 1, :]
T, U = mu_eta.shape
_, K = mu_c.shape
lpsi = np.einsum('u,tu,uk->tk', tau, mu_eta, mu_c)
lpsi += 0.5 * np.einsum('u,uk,uj,tj->tk', tau, mu_c, mu_c, xi1)
lpsi += 0.5 * np.einsum('u,ukj,tj->tk', tau, Sig_c, xi1)
lpsi += 0.5 * np.einsum('u,uk,uk,tk->tk', tau, mu_c, mu_c, 1 - xi1)
lpsi += 0.5 * np.einsum('u,ukk,tk->tk', tau, Sig_c, 1 - xi1)
log_psi = np.zeros((T, 2, K))
log_psi[:, 1, :] = lpsi
return log_psi
def _increment_negative_power_in_einsum_r(formula, x, exponent, args1, args2,
args3):
in_formulas, out_formula = split_einsum_formula(formula)
new_formula = _reconstitute_einsum_formula(
in_formulas[:len(args1) + 1 + len(args2)] +
in_formulas[len(args1) + 2 + len(args2):], out_formula)
return np.einsum(new_formula,
*(args1 + (x**(exponent + 1), ) + args2 + args3))
def _transpose_inside_einsum(formula, args1, x, args2):
in_formulas, out_formula = split_einsum_formula(formula)
i = len(args1)
new_formula = _reconstitute_einsum_formula(
in_formulas[:i] + [in_formulas[i][::-1]] + in_formulas[i + 1:],
out_formula)
new_args = args1 + (x, ) + args2
return np.einsum(new_formula, *new_args)
def hessian_local_log_likelihood(self, x):
"""
d/dx (y - lmbda)^T C = d/dx -exp(Cx + d)^T C
= -C^T exp(Cx + d)^T C
"""
# Observation likelihoods
lmbda = np.exp(np.dot(x, self.C.T) + np.dot(self.inputs, self.D.T))
return np.einsum('tn, ni, nj ->tij', -lmbda, self.C, self.C)
def dot(a, b):
if b.ndim == 1:
return _np.dot(a, b)
if a.ndim == 1:
return _np.dot(a, b.T)
return _np.einsum("...i,...i->...", a, b)
def mog_samples(N, means, chols, pis):
K, D = means.shape
indices = discrete(pis, (N,))
n_means = means[indices,:]
n_chols = chols[indices,:,:]
white = np.random.randn(N,D)
color = np.einsum('ikj,ij->ik', n_chols, white)
return color + n_means
def mixture_of_ts_samples(N, locs, scales, pis, df):
K, D = locs.shape
indices = mog.discrete(pis, (N, ))
n_means = locs[indices, :]
n_chols = scales[indices, :, :]
white = tdist.rvs(df=df, size=(N, D))
color = np.einsum('ikj,ij->ik', n_chols, white)
return color + n_means
def model(X, prior_precision):
beta = ph.norm.rvs(loc=0.0,
scale=1.0 / np.sqrt(prior_precision),
size=X.shape[1],
name="beta")
loc = np.einsum('ij,j->i', X, beta)
y = ph.norm.rvs(loc=loc, scale=1.0, name="y")
return y
def predictions(weights, inputs):
"""weights is shape (num_weight_samples x num_weights)
inputs is shape (num_datapoints x D)"""
inputs = np.expand_dims(inputs, 0)
for W, b in unpack_layers(weights):
outputs = np.einsum('mnd,mdo->mno', inputs, W) + b
inputs = nonlinearity(outputs)
return outputs
def _expand_integer_power_in_einsum(formula, x, exponent, args1, args2):
in_formulas, out_formula = split_einsum_formula(formula)
return np.einsum(
_reconstitute_einsum_formula(
in_formulas[:len(args1)] + [
in_formulas[len(args1)],
] * exponent + in_formulas[len(args1) + 1:], out_formula),
*(args1 + (x, ) * exponent + args2))
def bnn_predict(weights, inputs, layer_sizes, act):
if len(inputs.shape)<3: inputs = np.expand_dims(inputs, 0) # [1,N,D]
weights = reshape_weights(weights, layer_sizes)
for W, b in weights:
#print(W.shape, inputs.shape)
outputs = np.einsum('mnd,mdo->mno', inputs, W) + b
inputs = act_dict[act](outputs)
return outputs
def update(self, x, u, r, x_, u_):
rpe = r - self.uQx(u, x)
z = np.outer(u, x)
dQ = rpe * np.einsum('a,s->as', u, x)
if r >= 0:
self.Q += self.learning_rate_pos * dQ
else:
self.Q += self.learning_rate_neg * dQ
def _add_powers_within_einsum(formula, x, args1, args2, args3, exponent1,
exponent2):
in_formulas, out_formula = split_einsum_formula(formula)
new_formula = _reconstitute_einsum_formula(
_remove_list_elements(in_formulas, [len(args1) + 1 + len(args2)]),
out_formula)
return np.einsum(new_formula,
*(args1 + (x**(exponent1 + exponent2), ) + args2 + args3))
def predictions(weights, inputs):
"""weights is shape (num_weight_samples x num_weights)
inputs is shape (num_datapoints x D)"""
inputs = np.expand_dims(inputs, 0)
for W, b in unpack_layers(weights):
outputs = np.einsum('mnd,mdo->mno', inputs, W) + b
inputs = nonlinearity(outputs)
return outputs
def sufficientStats( cls, x, constParams=None ):
# Compute T( x )
x, y = x
assert isinstance( x, np.ndarray )
assert isinstance( y, np.ndarray )
if( x.ndim == 1 ):
# Only 1 point was passed in
x = x.reshape( ( 1, -1 ) )
assert y.ndim == 1 or y.ndim == 2
t2 = x.T.dot( x )
if( y.ndim == 1 ):
# 1 measurement for x
y = y.reshape( ( 1, -1 ) )
t1 = y.T.dot( y )
t3 = x.T.dot( y )
else:
# Multiple measurements for x
t2 *= y.shape[ 0 ]
t1 = np.einsum( 'mi,mj->ij', y, y )
t3 = np.einsum( 'i,mj->ij', x.ravel(), y )
else:
# Multiple data points were passed in
t2 = x.T.dot( x )
if( y.ndim == 3 ):
# Multiple measurements of y per x
assert x.shape[ 0 ] == y.shape[ 1 ]
t2 *= y.shape[ 0 ]
t1 = np.einsum( 'mti,mtj->ij', y, y )
t3 = np.einsum( 'ti,mtj->ij', x, y )
elif( y.ndim == 2 ):
# One measurement of y per x
assert x.shape[ 0 ] == y.shape[ 0 ]
t1 = y.T.dot( y )
t3 = x.T.dot( y )
else:
assert 0, 'Invalid dim'
# # For the sake of numerical precision
# t1 = ( t1 + t1.T ) / 2.0
# t2 = ( t2 + t2.T ) / 2.0
return t1, t2, t3
def _update_noderivatives(self, x, u, r, x_, u_):
""" Computes the value function update of the instrumental Rescorla-Wagner learning rule without the derivative.
This function is identical to `.update()` method except without the derivative computations. It is implemented solely for the purpose of unit testing the gradient calculations against `autograd`.
"""
rpe = r - self.uQx(u, x)
z = np.outer(u, x)
self.Q += self.learning_rate * rpe * np.einsum('a,s->as', u, x)
def logprob(z):
"""z is NxD."""
z_minus_mean = z - mean
if len(z.shape) == 1 or z.shape[0] == 1:
return const - 0.5 * np.dot(np.dot(z_minus_mean, pinv),
z_minus_mean.T)
else:
return const - 0.5 * np.einsum('ij,jk,ik->i', z_minus_mean, pinv,
z_minus_mean)
def mc_elbo(pgm_params, i):
#Here nn_potentials are just the sufficient stats of the data
x = get_batch(i)
xxT = np.einsum('ij,ik->ijk', x, x)
n = np.ones(x.shape[0]) if x.ndim == 2 else 1.
nn_potentials = pack_dense(xxT, x, n, n)
saved.stats, global_kl, local_kl = run_inference(
pgm_prior, pgm_params, nn_potentials)
return (-global_kl - num_batches * local_kl) / num_datapoints #CHECK
def log_joint(x, w, epsilon, tau, alpha, beta):
log_p_epsilon = log_probs.norm_gen_log_prob(epsilon, 0, 1)
log_p_w = log_probs.norm_gen_log_prob(w, 0, 1)
log_p_tau = log_probs.gamma_gen_log_prob(tau, alpha, beta)
# TODO(mhoffman): The transposed version below should work.
# log_p_x = log_probs.norm_gen_log_prob(x, np.dot(epsilon, w), 1. / np.sqrt(tau))
log_p_x = log_probs.norm_gen_log_prob(x, np.einsum('ik,jk->ij', epsilon, w),
1. / np.sqrt(tau))
return log_p_epsilon + log_p_w + log_p_tau + log_p_x
def nn_predict_tgcn_cheb(params, x):
L = graph.rescale_L(hyper['L'][0], lmax=2)
w = np.fft.fft(x, axis=2)
xc = chebyshev_time_vertex(L, w, hyper['filter_order'])
y = np.einsum('knhq,kfh->fnq', xc, params['W1'])
y += np.expand_dims(params['b1'], axis=2)
# nonlinear layer
# y = np.tanh(y)
y = ReLU(y)
# dense layer
y = np.einsum('fnq,cfn->cq', y, params['W2'])
y += np.expand_dims(params['b2'], axis=1)
outputs = np.real(y.T)
return outputs - logsumexp(outputs, axis=1, keepdims=True)
def outer(a, b):
if a.ndim == 2 and b.ndim == 2:
return _np.einsum("...i,...j->...ij", a, b)
out = _np.outer(a, b).reshape(a.shape + b.shape)
if b.ndim == 2:
out = out.swapaxes(-3, -2)
return out
def contextual_feature_map(self, features):
""" Creates contextual feature map
Args:
features (np.array): observation features
"""
features_squarred = np.einsum('ij,ih->ijh', features, features).reshape(features.shape[0], -1)
return np.hstack([features_squarred, features])
def tensorexp(r):
""" returns a stack of rotation matrices as a tensor """
""" r should be (3,n), n column vectors """
theta = np.sqrt(np.sum(r*r, axis=0)) # shape = (n,)
# note: the case where theta == 0 is not handled; we assume there is enough
# noise and bias that this won't happen
K = tensorhat(r / theta) # shape = (3,3,n)
KK = np.einsum('ijl,jkl->ikl', K, K)
# Compute w/ Rodrigues' formula
return np.eye(3)[:, :, np.newaxis] + np.sin(theta) * K + \
(1 - np.cos(theta)) * KK
def test_jacobian_against_wrapper():
A = npr.randn(3,3,3)
fun = lambda x: np.einsum(
'ijk,jkl->il',
A, np.sin(x[...,None] * np.tanh(x[None,...])))
B = npr.randn(3,3)
jac1 = jacobian(fun)(B)
jac2 = old_jacobian(fun)(B)
assert np.allclose(jac1, jac2)
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 make_pinwheel_data(radial_std, tangential_std, num_classes, num_per_class, rate):
rads = np.linspace(0, 2*np.pi, num_classes, endpoint=False)
features = npr.randn(num_classes*num_per_class, 2) \
* np.array([radial_std, tangential_std])
features[:,0] += 1.
labels = np.repeat(np.arange(num_classes), num_per_class)
angles = rads[labels] + rate * np.exp(features[:,0])
rotations = np.stack([np.cos(angles), -np.sin(angles), np.sin(angles), np.cos(angles)])
rotations = np.reshape(rotations.T, (-1, 2, 2))
return 10*npr.permutation(np.einsum('ti,tij->tj', features, rotations))
def pylds_E_step_inhomog(lds, data):
T = data.shape[0]
mu_init, sigma_init, A, sigma_states, C, sigma_obs = lds
normalizer, smoothed_mus, smoothed_sigmas, E_xtp1_xtT = \
_E_step(mu_init, sigma_init, A, sigma_states, C, sigma_obs, data)
EyyT = np.einsum('ti,tj->tij', data, data)
EyxT = np.einsum('ti,tj->tij', data, smoothed_mus)
ExxT = smoothed_sigmas + np.einsum('ti,tj->tij', smoothed_mus, smoothed_mus)
E_xt_xtT = ExxT[:-1]
E_xtp1_xtp1T = ExxT[1:]
E_xtp1_xtT = E_xtp1_xtT
E_x1_x1T = smoothed_sigmas[0] + np.outer(smoothed_mus[0], smoothed_mus[0])
E_x1 = smoothed_mus[0]
E_init_stats = E_x1_x1T, E_x1, 1.
E_pairwise_stats = E_xt_xtT.sum(0), E_xtp1_xtT.sum(0).T, E_xtp1_xtp1T.sum(0), T-1
E_node_stats = ExxT, np.transpose(EyxT, (0, 2, 1)), EyyT, np.ones(T)
return E_init_stats, E_pairwise_stats, E_node_stats
def get_next_layer(prev_layer, offset, A, Weights):
"""
Applies the Picard iteration to the current layer.
"""
N, _, _ = prev_layer.shape
# precompute A.dot(prev_layer)
Az = np.einsum('nij,njm->nim', A, prev_layer)
new_layer = []
for n in range(N):
new_layer.append(np.sum(Az * Weights[n, :][:, None, None], axis=0) + \
offset.T)
return np.array(new_layer)
def pair_mean_to_natural(A, sigma):
assert 2 <= A.ndim == sigma.ndim <= 3
ndim = A.ndim
einstring = 'tji,tjk->tik' if ndim == 3 else 'ji,jk->ik'
trans = (0, 2, 1) if ndim == 3 else (1, 0)
temp = np.linalg.solve(sigma, A)
Jxx = -1./2 * np.einsum(einstring, A, temp)
Jxy = np.transpose(temp, trans)
Jyy = -1./2 * np.linalg.inv(sigma)
logZ = -1./2 * np.linalg.slogdet(sigma)[1]
return Jxx, Jxy, Jyy, logZ
def make_pinwheel(radial_std, tangential_std, num_classes, num_per_class, rate,
rs=npr.RandomState(0)):
"""Based on code by Ryan P. Adams."""
rads = np.linspace(0, 2*np.pi, num_classes, endpoint=False)
features = rs.randn(num_classes*num_per_class, 2) \
* np.array([radial_std, tangential_std])
features[:, 0] += 1
labels = np.repeat(np.arange(num_classes), num_per_class)
angles = rads[labels] + rate * np.exp(features[:,0])
rotations = np.stack([np.cos(angles), -np.sin(angles), np.sin(angles), np.cos(angles)])
rotations = np.reshape(rotations.T, (-1, 2, 2))
return np.einsum('ti,tij->tj', features, rotations)
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 _ll(self, m, p, xn, **kwargs):
"""Computation of log likelihood
Dimensions
----------
m : n_unique x n_features
p : n_unique x n_features x n_features
xn: N x n_features
"""
samples = xn.shape[0]
xn = xn.reshape(samples, 1, self.n_features)
m = m.reshape(1, self.n_unique, self.n_features)
det = np.linalg.det(np.linalg.inv(p))
det = det.reshape(1, self.n_unique)
tem = np.einsum('NUF,UFX,NUX->NU', (xn - m), p, (xn - m))
res = (-self.n_features/2.0)*np.log(2*np.pi) - 0.5*tem - 0.5*np.log(det)
return res # N x n_unique
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 edges_score(u):
U = anp.reshape(u, (n, d))
m = U[edges[:, 0]] * U[edges[:, 1]]
return anp.einsum('ij,ji->i', m, ((1 + mu / k) * W[wc, :].T - vecc[:, np.newaxis])).mean()