def initialize_from_gaussian_lds(self, N_samples=100):
"""
Initialize z, A, C, sigma_states using a Gaussian LDS
:return:
"""
from pylds.models import DefaultLDS
init_model = DefaultLDS(n=self.n, p=self.K)
for data in self.data_list:
init_model.add_data(data["x"])
print("Initializing with Gaussian LDS")
for smpl in range(20):
init_model.resample_model()
# Use the init model's parameters
self.A = init_model.A.copy()
self.C = init_model.C[:self.K-1,:].copy()
self.sigma_states = init_model.sigma_states.copy()
self.mu_init = init_model.mu_init.copy()
self.sigma_init = init_model.sigma_init.copy()
# Use the init model's latent state sequences too
for data, init_data in zip(self.data_list, init_model.states_list):
data["z"] = init_data.stateseq.copy()
# Now resample omega
self.emission_distn.resample_omega(self.data_list)
def random_model(n, p, T):
data = np.random.randn(T, p)
model = DefaultLDS(n, p)
model.A = 0.99 * random_rotation(n, 0.01)
model.C = np.random.randn(p, n)
J, h = lds_to_dense_infoparams(model, data)
model.add_data(data)
return model, (J, h)
def fit_lds_gibbs(seq, inputs, guessed_dim, num_update_samples):
"""Fits LDS model via Gibbs sampling and EM. Returns fitted eigenvalues."""
if inputs is None:
model = DefaultLDS(D_obs=1, D_latent=guessed_dim, D_input=0)
else:
model = DefaultLDS(D_obs=1, D_latent=guessed_dim, D_input=1)
model.add_data(seq, inputs=inputs)
ll = np.zeros(num_update_samples)
# Run the Gibbs sampler
for i in xrange(num_update_samples):
try:
model.resample_model()
except AssertionError as e:
warnings.warn(str(e), sm_exceptions.ConvergenceWarning)
eigs = np.linalg.eigvals(model.A)
return eigs[np.argsort(np.abs(eigs))[::-1]]
ll[i] = model.log_likelihood()
# Rough estimate of convergence: judge converged if the change of maximum
# log likelihood is less than tolerance.
recent_steps = int(num_update_samples / 10)
tol = 1.0
if np.max(ll[-recent_steps:]) - np.max(ll[:-recent_steps]) > tol:
warnings.warn(
'Questionable convergence. Log likelihood values: ' + str(ll),
sm_exceptions.ConvergenceWarning)
eigs = np.linalg.eigvals(model.A)
return eigs[np.argsort(eigs.real)[::-1]]
def fit_gaussian_lds_model(Xs, Xtest, D_gauss_lds, N_samples=100):
print("Fitting Gaussian (Raw) LDS with %d states" % D_gauss_lds)
model = DefaultLDS(n=D_gauss_lds, p=K)
Xs_centered = [X - np.mean(X, axis=0)[None,:] + 1e-3*np.random.randn(*X.shape) for X in Xs]
for X in Xs_centered:
model.add_data(X)
# TODO: Get initial pred ll
init_results = (0, None, np.nan, np.nan, np.nan)
def resample():
tic = time.time()
model.resample_model()
toc = time.time() - tic
# Monte Carlo sample to get pi density implied by Gaussian LDS
Tpred = Xtest.shape[0]
Npred = 1000
preds = model.sample_predictions(Xs_centered[0], Tpred, Npred=Npred)
# Convert predictions to a distribution by finding the
# largest dimension for each predicted Gaussian.
# Preds is T x K x Npred, inds is TxNpred
inds = np.argmax(preds, axis=1)
pi = np.array([np.bincount(inds[t], minlength=K) for t in xrange(Tpred)]) / float(Npred)
assert np.allclose(pi.sum(axis=1), 1.0)
pi = np.clip(pi, 1e-8, 1.0)
pi /= pi.sum(axis=1)[:,None]
# Compute the log likelihood under pi
pred_ll = np.sum([Multinomial(weights=pi[t], K=K).log_likelihood(Xtest[t][None,:])
for t in xrange(Tpred)])
return toc, None, np.nan, \
np.nan, \
pred_ll
n_retries = 0
max_attempts = 5
while n_retries < max_attempts:
try:
times, samples, lls, test_lls, pred_lls = \
map(np.array, zip(*([init_results] + [resample() for _ in progprint_xrange(N_samples)])))
timestamps = np.cumsum(times)
return Results(lls, test_lls, pred_lls, samples, timestamps)
except Exception as e:
print("Caught exception: ", e.message)
print("Retrying")
n_retries += 1
raise Exception("Failed to fit the Raw Gaussian LDS model in %d attempts" % max_attempts)
def random_model(n, p, d, T):
data = np.random.randn(T, p)
inputs = np.random.randn(T, d)
model = DefaultLDS(p, n, d)
model.A = 0.99 * random_rotation(n, 0.01)
model.B = 0.1 * np.random.randn(n, d)
model.C = np.random.randn(p, n)
model.D = 0.1 * np.random.randn(p, d)
J, h = lds_to_dense_infoparams(model, data, inputs)
model.add_data(data, inputs=inputs)
return model, (J, h)
def initialize_from_gaussian_lds(self, N_samples=100):
"""
Initialize z, A, C, sigma_states using a Gaussian LDS
:return:
"""
from pylds.models import DefaultLDS
init_model = DefaultLDS(n=self.n, p=self.K)
for data in self.data_list:
init_model.add_data(data["x"])
print("Initializing with Gaussian LDS")
for smpl in range(20):
init_model.resample_model()
# Use the init model's parameters
self.A = init_model.A.copy()
self.C = init_model.C[:self.K - 1, :].copy()
self.sigma_states = init_model.sigma_states.copy()
self.mu_init = init_model.mu_init.copy()
self.sigma_init = init_model.sigma_init.copy()
# Use the init model's latent state sequences too
for data, init_data in zip(self.data_list, init_model.states_list):
data["z"] = init_data.stateseq.copy()
# Now resample omega
self.emission_distn.resample_omega(self.data_list)
def fit_gaussian_lds_model(Xs, N_samples=100):
testmodel = DefaultLDS(n=D, p=K)
for X in Xs:
testmodel.add_data(X)
samples = []
lls = []
for smpl in progprint_xrange(N_samples):
testmodel.resample_model()
samples.append(testmodel.copy_sample())
lls.append(testmodel.log_likelihood())
lls = np.array(lls)
return lls
def random_model(n, p, T):
data = np.random.randn(T, p)
model = DefaultLDS(n, p)
model.A = 0.99 * random_rotation(n, 0.01)
model.C = np.random.randn(p, n)
J, h = lds_to_dense_infoparams(model, data)
model.add_data(data)
return model, (J, h)
def fit_gaussian_lds_model(Xs, N_samples=100):
testmodel = DefaultLDS(n=D,p=K)
for X in Xs:
testmodel.add_data(X)
samples = []
lls = []
for smpl in progprint_xrange(N_samples):
testmodel.resample_model()
samples.append(testmodel.copy_sample())
lls.append(testmodel.log_likelihood())
lls = np.array(lls)
return lls
from pybasicbayes.distributions import Regression, DiagonalRegression
from pybasicbayes.util.text import progprint_xrange
from pylds.models import LDS, DefaultLDS
npr.seed(0)
# Parameters
D_obs = 1
D_latent = 2
D_input = 0
T = 2000
# Simulate from an LDS with diagonal observation noise
truemodel = DefaultLDS(D_obs, D_latent, D_input, sigma_obs=0.1 * np.eye(D_obs))
inputs = np.random.randn(T, D_input)
data, stateseq = truemodel.generate(T, inputs=inputs)
# Fit with an LDS with diagonal observation noise
diag_model = LDS(dynamics_distn=Regression(
nu_0=D_latent + 2,
S_0=D_latent * np.eye(D_latent),
M_0=np.zeros((D_latent, D_latent + D_input)),
K_0=(D_latent + D_input) * np.eye(D_latent + D_input)),
emission_distn=DiagonalRegression(D_obs, D_latent + D_input))
diag_model.add_data(data, inputs=inputs)
# Also fit a model with a full covariance matrix
full_model = DefaultLDS(D_obs, D_latent, D_input)
full_model.add_data(data, inputs=inputs)
import matplotlib.pyplot as plt
from pybasicbayes.distributions import Regression, DiagonalRegression
from pybasicbayes.util.text import progprint_xrange
from pylds.models import DefaultLDS, MissingDataLDS
npr.seed(0)
# Model parameters
D_obs = 4
D_latent = 4
T = 1000
# Simulate from an LDS
truemodel = DefaultLDS(D_obs, D_latent)
data, stateseq = truemodel.generate(T)
# Mask off a chunk of data
mask = np.ones_like(data, dtype=bool)
chunksz = 100
for i,offset in enumerate(range(0,T,chunksz)):
j = i % (D_obs + 1)
if j < D_obs:
mask[offset:min(offset+chunksz, T), j] = False
if j == D_obs:
mask[offset:min(offset+chunksz, T), :] = False
# Fit with another LDS
model = MissingDataLDS(
dynamics_distn=Regression(
def fit_gaussian_lds_model(Xs, Xtest, D_gauss_lds, N_samples=100):
print("Fitting Gaussian (Raw) LDS with %d states" % D_gauss_lds)
model = DefaultLDS(n=D_gauss_lds, p=K)
Xs_centered = [
X - np.mean(X, axis=0)[None, :] + 1e-3 * np.random.randn(*X.shape)
for X in Xs
]
for X in Xs_centered:
model.add_data(X)
# TODO: Get initial pred ll
init_results = (0, None, np.nan, np.nan, np.nan)
def resample():
tic = time.time()
model.resample_model()
toc = time.time() - tic
# Monte Carlo sample to get pi density implied by Gaussian LDS
Tpred = Xtest.shape[0]
Npred = 1000
preds = model.sample_predictions(Xs_centered[0], Tpred, Npred=Npred)
# Convert predictions to a distribution by finding the
# largest dimension for each predicted Gaussian.
# Preds is T x K x Npred, inds is TxNpred
inds = np.argmax(preds, axis=1)
pi = np.array(
[np.bincount(inds[t], minlength=K)
for t in xrange(Tpred)]) / float(Npred)
assert np.allclose(pi.sum(axis=1), 1.0)
pi = np.clip(pi, 1e-8, 1.0)
pi /= pi.sum(axis=1)[:, None]
# Compute the log likelihood under pi
pred_ll = np.sum([
Multinomial(weights=pi[t], K=K).log_likelihood(Xtest[t][None, :])
for t in xrange(Tpred)
])
return toc, None, np.nan, \
np.nan, \
pred_ll
n_retries = 0
max_attempts = 5
while n_retries < max_attempts:
try:
times, samples, lls, test_lls, pred_lls = \
map(np.array, zip(*([init_results] + [resample() for _ in progprint_xrange(N_samples)])))
timestamps = np.cumsum(times)
return Results(lls, test_lls, pred_lls, samples, timestamps)
except Exception as e:
print("Caught exception: ", e.message)
print("Retrying")
n_retries += 1
raise Exception("Failed to fit the Raw Gaussian LDS model in %d attempts" %
max_attempts)
def createExactPongData(T, D_obs, Start, Per, Amp):
exactData = np.empty((T, D_obs))
exactData[0] = [Start] * D_obs
for t in range(1, T):
exactData[t] = calcDotPosition(t, Start, Per, Amp)
return exactData
data = createExactPongData(T, D_obs, Start, Per, Amp)
#pp.pprint(data)
# Fit with another LDS
model = DefaultLDS(D_obs, D_latent)
model.add_data(data)
# Initialize with a few iterations of Gibbs
for _ in progprint_xrange(10):
model.resample_model()
# Run EM
def update(model):
vlb = model.meanfield_coordinate_descent_step()
return vlb
vlbs = [update(model) for _ in progprint_xrange(50)]
num_mc_samples = 10
"""Model selection / Hypothesis tesing"""
# Now calculate
# log posterior_odds = log prior_odds + log likelihood_ratio
# - log posterior_odds = log p(H1|DI)/P(H2|DI)
# - log prior_odds = log P(H1|I)/P(H2I)
# - log likelihood_ratio = log P(D|H1I)/P(D|H2I)
# The variable I represents all our background information
# Let's assume our prior belief in both hypotheses is equal: P(H1|I) = P(H2I)
# Now log_prior_odds is then log(1) = 0
log_prior_odds = 0
# Calculate log P(D|HI) by integrating out theta in p(Dtheta|HI)=p(D|thetaHI)p(theta|HI)
print('Hypothesis 1')
model = DefaultLDS(D_obs=1, D_latent=2)
log_p_D_given_H1I = []
for _ in range(num_mc_samples):
model.resample_parameters()
log_p_D_given_H1I.append(
np.sum([
model.log_likelihood(np.expand_dims(data[n], 1))
for n in range(num_samples)
]))
# In the next line, we do a log-sum-exp over our list.
# - The outer log puts the evidence on log scale
# - The sum is over the MC samples
# - The exp cancels the log in the distribution.logpdf()
log_p_D_given_H1I = logsumexp(log_p_D_given_H1I) - np.log(num_mc_samples)
truemodel = LDS(dynamics_distn=AutoRegression(A=A, sigma=sigma_states),
emission_distn=Regression(A=C, sigma=sigma_obs))
data, stateseq = truemodel.generate(2000)
###############
# fit model #
###############
def update(model):
model.resample_model()
return model.log_likelihood()
model = DefaultLDS(n=2, p=data.shape[1]).add_data(data)
vlbs = [update(model) for _ in progprint_xrange(100)]
plt.figure(figsize=(3, 4))
plt.plot(vlbs)
plt.xlabel('iteration')
plt.ylabel('variational lower bound')
################
# predicting #
################
Npredict = 100
prediction_seed = data[:1700]
predictions = model.sample_predictions(prediction_seed,
truemodel = LDS(
dynamics_distn=AutoRegression(A=A,sigma=sigma_states),
emission_distn=Regression(A=C,sigma=sigma_obs))
data, stateseq = truemodel.generate(2000)
###############
# fit model #
###############
def update(model):
return model.meanfield_coordinate_descent_step()
model = DefaultLDS(n=2,p=data.shape[1]).add_data(data)
for _ in progprint_xrange(100):
model.resample_model()
vlbs = [update(model) for _ in progprint_xrange(50)]
plt.figure(figsize=(3,4))
plt.plot(vlbs)
plt.xlabel('iteration')
plt.ylabel('variational lower bound')
################
# predicting #
################
C = np.array([[10.,0.]])
sigma_obs = 0.01*np.eye(1)
# C = np.eye(2)
# sigma_obs = 0.01*np.eye(2)
###################
# generate data #
###################
truemodel = LDS(
dynamics_distn=AutoRegression(A=A,sigma=sigma_states),
emission_distn=Regression(A=C,sigma=sigma_obs))
data, stateseq = truemodel.generate(2000)
###############
# fit model #
###############
model = DefaultLDS(n=2,p=data.shape[1]).add_data(data)
likes = []
for _ in progprint_xrange(50):
model.EM_step()
likes.append(model.log_likelihood())
plt.plot(likes)
plt.show()
dynamics_distn=AutoRegression(A=A,sigma=sigma_states),
emission_distn=Regression(A=C,sigma=sigma_obs))
data, stateseq = truemodel.generate(2000)
###############
# fit model #
###############
def update(model):
model.EM_step()
return model.log_likelihood()
model = DefaultLDS(n=2,p=data.shape[1]).add_data(data)
likes = [update(model) for _ in progprint_xrange(50)]
plt.figure(figsize=(3,4))
plt.plot(likes)
plt.xlabel('iteration')
plt.ylabel('training likelihood')
################
# predicting #
################
Npredict = 100
prediction_seed = data[:1700]
from pybasicbayes.distributions import Regression, DiagonalRegression
from pybasicbayes.util.text import progprint_xrange
from pylds.models import LDS, DefaultLDS
npr.seed(0)
# Parameters
D_obs = 1
D_latent = 2
D_input = 0
T = 2000
# Simulate from an LDS
truemodel = DefaultLDS(D_obs, D_latent, D_input)
inputs = np.random.randn(T, D_input)
data, stateseq = truemodel.generate(T, inputs=inputs)
# Fit with an LDS with diagonal observation noise
model = LDS(dynamics_distn=Regression(
nu_0=D_latent + 2,
S_0=D_latent * np.eye(D_latent),
M_0=np.zeros((D_latent, D_latent + D_input)),
K_0=(D_latent + D_input) * np.eye(D_latent + D_input)),
emission_distn=DiagonalRegression(D_obs, D_latent + D_input))
model.add_data(data, inputs=inputs)
# Fit with mean field
def update(model):
import matplotlib.pyplot as plt
from pybasicbayes.distributions import Regression, DiagonalRegression
from pybasicbayes.util.text import progprint_xrange
from pylds.models import DefaultLDS, LDS
npr.seed(0)
# Model parameters
D_obs = 4
D_latent = 4
T = 1000
# Simulate from an LDS
truemodel = DefaultLDS(D_obs, D_latent)
data, stateseq = truemodel.generate(T)
# Mask off a chunk of data
mask = np.ones_like(data, dtype=bool)
chunksz = 100
for i, offset in enumerate(range(0, T, chunksz)):
j = i % (D_obs + 1)
if j < D_obs:
mask[offset:min(offset + chunksz, T), j] = False
if j == D_obs:
mask[offset:min(offset + chunksz, T), :] = False
# Fit with another LDS
model = LDS(dynamics_distn=Regression(nu_0=D_latent + 3,
S_0=D_latent * np.eye(D_latent),
import matplotlib.pyplot as plt
from pybasicbayes.util.text import progprint_xrange
from pylds.models import DefaultLDS
npr.seed(0)
# Set parameters
D_obs = 1
D_latent = 2
D_input = 0
T = 2000
# Simulate from one LDS
truemodel = DefaultLDS(D_obs, D_latent, D_input)
inputs = np.random.randn(T, D_input)
data, stateseq = truemodel.generate(T, inputs=inputs)
# Fit with another LDS
model = DefaultLDS(D_obs, D_latent, D_input)
model.add_data(data, inputs=inputs)
# Initialize with a few iterations of Gibbs
for _ in progprint_xrange(10):
model.resample_model()
# Run EM
def update(model):
model.EM_step()
except:
colors = ['b', 'r', 'y', 'g']
from pybasicbayes.util.text import progprint_xrange
from pylds.models import DefaultLDS
npr.seed(3)
# Set parameters
D_obs = 1
D_latent = 2
D_input = 0
T = 2000
# Simulate from one LDS
true_model = DefaultLDS(D_obs, D_latent, D_input, sigma_obs=np.eye(D_obs))
inputs = npr.randn(T, D_input)
data, stateseq = true_model.generate(T, inputs=inputs)
# Fit with another LDS
test_model = DefaultLDS(D_obs, D_latent, D_input)
test_model.add_data(data, inputs=inputs)
# Run the Gibbs sampler
N_samples = 100
def update(model):
model.resample_model()
return model.log_likelihood()
