def _start(self):
print("model")
print(self.model_dict)
n_features = float(self.model_dict['n_features'])
n_states = float(self.model_dict['n_states'])
self.model = MetastableSwitchingLDS(n_states, n_features)
self.model.load_from_json_dict(self.model_dict)
obs, hidden_states = self.model.sample(self.args.n_samples)
(n_samples, n_features) = np.shape(obs)
features, ii, ff = mixtape.featurizer.featurize_all(
self.filenames, self.featurizer, self.topology, self.stride)
file_trajectories = []
states = []
state_indices = []
state_files = []
logprob = log_multivariate_normal_density(
features, np.array(self.model.means_),
np.array(self.model.covars_), covariance_type='full')
assignments = np.argmax(logprob, axis=1)
probs = np.max(logprob, axis=1)
# Presort the data into the metastable wells
# i.e.: separate the original trajectories into k
# buckets corresponding to the metastable wells
for k in range(int(self.model.n_states)):
# pick the structures that have the highest log
# probability in the state
s = features[assignments == k]
ind = ii[assignments==k]
f = ff[assignments==k]
states.append(s)
state_indices.append(ind)
state_files.append(f)
# Loop over the generated feature space trajectory.
# At time t, pick the frame from the original trajectory
# closest to the current sample in feature space. To save
# a bit of computation, just search in the bucket corresponding
# to the current metastable well (i.e., the current hidden state).
traj = None
for t in range(n_samples):
featurized_frame = obs[t]
h = hidden_states[t]
logprob = log_multivariate_normal_density(
states[h], featurized_frame[np.newaxis],
self.model.Qs_[h][np.newaxis],
covariance_type='full')
best_frame_pos = np.argmax(logprob, axis=0)[0]
best_file = state_files[h][best_frame_pos]
best_ind = state_indices[h][best_frame_pos]
frame = md.load_frame(best_file, best_ind, self.topology)
if t == 0:
traj = frame
else:
frame.superpose(traj, t-1)
traj = traj.join(frame)
traj.save('%s.xtc' % self.out)
traj[0].save('%s.xtc.pdb' % self.out)
def _exact_loglikelihood(self, ob):
log_transmat = np.zeros((self.n_chains, self.n_states, self.n_states))
log_startprob = np.zeros((self.n_chains, self.n_states))
for idx, chain in enumerate(self.chains_):
log_transmat[idx] = chain._log_transmat
log_startprob[idx] = chain._log_startprob
n_state_combinations = self.n_states ** self.n_chains
state_combinations = [tuple(x) for x in list(itertools.product(np.arange(self.n_states), repeat=self.n_chains))]
n_observations = ob.shape[0]
n_features = ob.shape[1]
fwdlattice = np.zeros((n_observations, n_state_combinations))
# Calculate means and covariances for all state combinations and calculate emission probabilities
weight = (1.0 / float(self.n_chains))
weight_squared = weight * weight
covars = np.zeros((n_state_combinations, n_features)) # TODO: add support for all covariance types
means = np.zeros((n_state_combinations, n_features))
for idx, state_combination in enumerate(state_combinations):
for chain_idx, state in enumerate(state_combination):
chain = self.chains_[chain_idx]
covars[idx] += chain._covars_[state]
means[idx] += chain._means_[state]
covars[idx] *= weight_squared
means[idx] *= weight
framelogprob = log_multivariate_normal_density(ob, means, covars, covariance_type='diag') # TODO: add support for all covariance types
# Run the forward algorithm
fhmmc._forward(n_observations, self.n_chains, self.n_states, state_combinations, log_startprob, log_transmat,
framelogprob, fwdlattice)
last_column = fwdlattice[-1]
assert np.size(last_column) == n_state_combinations
score = logsumexp(last_column)
return score
def start(self):
featurizer = mixtape.featurizer.load(self.args.featurizer)
features, ii, ff = mixtape.featurizer.featurize_all(
self.filenames, featurizer, self.topology, self.stride)
logprob = log_multivariate_normal_density(
features, np.array(self.model['means']),
np.array(self.model['vars']), covariance_type='diag')
assignments = np.argmax(logprob, axis=1)
probs = np.max(logprob, axis=1)
data = {'filename': [], 'index': [], 'state': []}
for k in range(self.model['n_states']):
# pick the structures that have the highest log
# probability in the state
p = probs[assignments == k]
sorted_filenms = ff[assignments == k][p.argsort()]
sorted_indices = ii[assignments == k][p.argsort()]
if len(p) > 0:
data['index'].extend(sorted_indices[-self.args.n_per_state:])
index_length = len(sorted_indices[-self.args.n_per_state:])
data['filename'].extend(sorted_filenms[-self.args.n_per_state:])
filename_length = len(sorted_filenms[-self.args.n_per_state:])
assert index_length == filename_length
data['state'].extend([k] * index_length)
else:
print('WARNING: NO STRUCTURES ASSIGNED TO STATE=%d' % k)
df = pd.DataFrame(data)
print('Saving the indices of the selected frames in CSV format to %s' % self.out)
with open(self.out, 'w') as f:
f.write("# command: %s\n" % ' '.join(sys.argv))
df.to_csv(f)
def get_word_ll_for_topics(self): """Calculates P(w|z) for all words (rows), given each topic (columns) Returns matrix that is VxK (where V is the number of words in the vocabulary, and K is the number of topics) """ word_vectors = self._w2v_model.syn0 return log_multivariate_normal_density(word_vectors, self._gmm_model.means_, self._gmm_model.covars_, self._gmm_model.covariance_type)
def score(self, X, return_responsibilities=False):
nc = len(self.weights)
X = numpy.array(X)
if X.ndim == 1:
X = X[:, None]
if X.shape[1] != self.means.shape[1]:
raise ValueError('The shape of X is not compatible with self')
mins = self.lims[:, 0]
maxes = self.lims[:, 1]
lpr = numpy.log(self.weights) + \
log_multivariate_normal_density(X,
self.means,
self.covs, 'full')
mask = (X >= mins[None, :]).all(axis=-1)
mask &= (X <= maxes[None, :]).all(axis=-1)
logprob = logsumexp(lpr, axis=1)
logprob[~mask] = -numpy.inf
if return_responsibilities:
responsibilities = numpy.exp(lpr - logprob[:, None])
responsibilities[~mask] = 0
return logprob, responsibilities
return logprob
def get_word_ll_for_topics(self):
"""Calculates P(w|z) for all words (rows), given each topic (columns)
Returns matrix that is VxK (where V is the number of words in the vocabulary, and K is the number of topics)
"""
word_vectors = self._w2v_model.syn0
return log_multivariate_normal_density(word_vectors,
self._gmm_model.means_,
self._gmm_model.covars_,
self._gmm_model.covariance_type)
def _compute_log_likelihood(self, seq):
state_combinations = [tuple(x) for x in list(itertools.product(np.arange(self.n_states), repeat=self.n_chains))]
n_state_combinations = self.n_states ** self.n_chains
n_observations, n_features = seq.shape
covars = np.array([self.covar for _ in xrange(n_state_combinations)]) # TODO: correct?!
means = np.zeros((n_state_combinations, n_features))
for idx, state_combination in enumerate(state_combinations):
for chain_idx, state in enumerate(state_combination):
means[idx] += self.means[chain_idx, state]
framelogprob = log_multivariate_normal_density(seq, means, covars, covariance_type='full')
return framelogprob
def test_framelogprob_reshape(self):
n_states = 3
n_chains = 2
n_state_combinations = n_states ** n_chains
state_combinations = [tuple(x) for x in list(itertools.product(np.arange(n_states), repeat=n_chains))]
covars = np.random.random((n_state_combinations, 10))
means = np.random.random((n_state_combinations, 10))
ob = np.random.random((5, 10))
framelogprob = log_multivariate_normal_density(ob, means, covars, covariance_type='diag')
# This test assures that resizing the framelogprob still yields the correct state variables
reshaped_framelogprob = framelogprob.reshape((5, n_states, n_states))
for ob_idx in xrange(5):
for idx, state_combination in enumerate(state_combinations):
self.assertEqual(reshaped_framelogprob[ob_idx][state_combination], framelogprob[ob_idx][idx])
def _exact_loglikelihood(self, ob):
log_transmat = np.zeros((self.n_chains, self.n_states, self.n_states))
log_startprob = np.zeros((self.n_chains, self.n_states))
for idx, chain in enumerate(self.chains_):
log_transmat[idx] = chain._log_transmat
log_startprob[idx] = chain._log_startprob
n_state_combinations = self.n_states**self.n_chains
state_combinations = [
tuple(x) for x in list(
itertools.product(np.arange(self.n_states),
repeat=self.n_chains))
]
n_observations = ob.shape[0]
n_features = ob.shape[1]
fwdlattice = np.zeros((n_observations, n_state_combinations))
# Calculate means and covariances for all state combinations and calculate emission probabilities
weight = (1.0 / float(self.n_chains))
weight_squared = weight * weight
covars = np.zeros(
(n_state_combinations,
n_features)) # TODO: add support for all covariance types
means = np.zeros((n_state_combinations, n_features))
for idx, state_combination in enumerate(state_combinations):
for chain_idx, state in enumerate(state_combination):
chain = self.chains_[chain_idx]
covars[idx] += chain._covars_[state]
means[idx] += chain._means_[state]
covars[idx] *= weight_squared
means[idx] *= weight
framelogprob = log_multivariate_normal_density(
ob, means, covars, covariance_type='diag'
) # TODO: add support for all covariance types
# Run the forward algorithm
fhmmc._forward(n_observations, self.n_chains, self.n_states,
state_combinations, log_startprob, log_transmat,
framelogprob, fwdlattice)
last_column = fwdlattice[-1]
assert np.size(last_column) == n_state_combinations
score = logsumexp(last_column)
return score
def _compute_log_likelihood(self, seq):
state_combinations = [
tuple(x) for x in list(
itertools.product(np.arange(self.n_states),
repeat=self.n_chains))
]
n_state_combinations = self.n_states**self.n_chains
n_observations, n_features = seq.shape
covars = np.array([self.covar for _ in xrange(n_state_combinations)
]) # TODO: correct?!
means = np.zeros((n_state_combinations, n_features))
for idx, state_combination in enumerate(state_combinations):
for chain_idx, state in enumerate(state_combination):
means[idx] += self.means[chain_idx, state]
framelogprob = log_multivariate_normal_density(seq,
means,
covars,
covariance_type='full')
return framelogprob
def start(self):
featurizer = mixtape.featurizer.load(self.args.featurizer)
features, ii, ff = mixtape.featurizer.featurize_all(
self.filenames, featurizer, self.topology)
logprob = log_multivariate_normal_density(features,
np.array(
self.model['means']),
np.array(self.model['vars']),
covariance_type='diag')
assignments = np.argmax(logprob, axis=1)
probs = np.max(logprob, axis=1)
data = {'filename': [], 'index': [], 'state': []}
for k in range(self.model['n_states']):
# pick the structures that have the highest log
# probability in the state
p = probs[assignments == k]
sorted_filenms = ff[assignments == k][p.argsort()]
sorted_indices = ii[assignments == k][p.argsort()]
if len(p) > 0:
data['index'].extend(sorted_indices[-self.args.n_per_state:])
data['filename'].extend(
sorted_filenms[-self.args.n_per_state:])
data['state'].extend([k] * self.args.n_per_state)
else:
print('WARNING: NO STRUCTURES ASSIGNED TO STATE=%d' % k)
df = pd.DataFrame(data)
print('Saving the indices of the selected frames in CSV format to %s' %
self.out)
with open(self.out, 'w') as f:
f.write("# command: %s\n" % ' '.join(sys.argv))
df.to_csv(f)
def _compute_log_likelihood(self, X):
"""Compute the log likelihood of feature matrix X"""
return gmm.log_multivariate_normal_density(X, self.means_,
self.covars_,
self.covariance_type)
def test_alanine_dipeptide():
import pdb, traceback, sys
warnings.filterwarnings("ignore",
category=DeprecationWarning)
try:
b = fetch_alanine_dipeptide()
trajs = b.trajectories
n_seq = len(trajs)
n_frames = trajs[0].n_frames
n_atoms = trajs[0].n_atoms
n_features = n_atoms * 3
sim_T = 1000
data_home = get_data_home()
data_dir = join(data_home, TARGET_DIRECTORY_ALANINE)
top = md.load(join(data_dir, 'ala2.pdb'))
n_components = 2
# Superpose m
data = []
for traj in trajs:
traj.superpose(top)
Z = traj.xyz
Z = np.reshape(Z, (len(Z), n_features), order='F')
data.append(Z)
# Fit MSLDS model
n_experiments = 1
n_em_iter = 1
tol = 1e-1
model = MetastableSwitchingLDS(n_components,
n_features, n_experiments=n_experiments,
n_em_iter=n_em_iter)
model.fit(data, gamma=.1, tol=tol, verbose=True)
mslds_score = model.score(data)
print("MSLDS Log-Likelihood = %f" % mslds_score)
# Fit Gaussian HMM for comparison
g = GaussianFusionHMM(n_components, n_features)
g.fit(data)
hmm_score = g.score(data)
print("HMM Log-Likelihood = %f" % hmm_score)
print()
# Generate a trajectory from learned model.
sample_traj, hidden_states = model.sample(sim_T)
states = []
for k in range(n_components):
states.append([])
# Presort the data into the metastable wells
for k in range(n_components):
for i in range(len(trajs)):
traj = trajs[i]
Z = traj.xyz
Z = np.reshape(Z, (len(Z), n_features), order='F')
logprob = log_multivariate_normal_density(Z,
np.array(model.means_),
np.array(model.covars_), covariance_type='full')
assignments = np.argmax(logprob, axis=1)
#probs = np.max(logprob, axis=1)
# pick structures that have highest log probability in state
s = traj[assignments == k]
states[k].append(s)
# Pick frame from original trajectories closest to current sample
gen_traj = None
for t in range(sim_T):
h = hidden_states[t]
best_logprob = -np.inf
best_frame = None
for i in range(len(trajs)):
if t > 0:
states[h][i].superpose(gen_traj, t-1)
Z = states[h][i].xyz
Z = np.reshape(Z, (len(Z), n_features), order='F')
mean = sample_traj[t]
logprobs = log_multivariate_normal_density(Z,
mean, model.Qs_[h], covariance_type='full')
ind = np.argmax(logprobs, axis=0)
logprob = logprobs[ind]
if logprob > best_log_prob:
logprob = best_logprob
best_frame = states[h][i][ind]
if t == 0:
gen_traj = best_frame
else:
gen_traj = gen_traj.join(frame)
gen_traj.save('%s.xtc' % self.out)
gen_traj[0].save('%s.xtc.pdb' % self.out)
except:
type, value, tb = sys.exc_info()
traceback.print_exc()
pdb.post_mortem(tb)
# Train GMM
print 'Starting GMM training'
words = w2v_model.vocab.keys()
word_vectors = w2v_model.syn0
gmm_model = GMM(n_components=num_topics, n_iter=num_gmm_iterations, covariance_type='diag')
gmm_model.fit(word_vectors)
# joblib.dump(gmm_model, gmm_output_file)
print 'Done GMM training'
# Get the likelihood of each word vector under each Gaussian component
scores = gmm_model.score(test_vectors)
print scores
ll = sum(scores)
print "LL: "+str(ll)
# Print topics if desired
if print_topics:
log_probs = log_multivariate_normal_density(word_vectors, gmm_model.means_, gmm_model.covars_, gmm_model.covariance_type)
print np.min(log_probs)
_, num_col = log_probs.shape
for col in xrange(num_col):
top_n = 10
log_component_probs = (log_probs[:,col]).T
sorted_indexes = np.argsort(log_component_probs)[::-1][:top_n]
ordered_word_probs = [(w2v_model.index2word[idx], log_component_probs[idx]) for idx in sorted_indexes]
print '---'
print "Topic {0}".format(col+1)
print "Total prob:" + str(sum(log_component_probs))
print ", ".join(["{w}: {p}".format(w=w, p=p) for w, p in ordered_word_probs])
n_iter=num_gmm_iterations,
covariance_type='diag')
gmm_model.fit(word_vectors)
# joblib.dump(gmm_model, gmm_output_file)
print 'Done GMM training'
# Get the likelihood of each word vector under each Gaussian component
scores = gmm_model.score(test_vectors)
print scores
ll = sum(scores)
print "LL: " + str(ll)
# Print topics if desired
if print_topics:
log_probs = log_multivariate_normal_density(
word_vectors, gmm_model.means_, gmm_model.covars_,
gmm_model.covariance_type)
print np.min(log_probs)
_, num_col = log_probs.shape
for col in xrange(num_col):
top_n = 10
log_component_probs = (log_probs[:, col]).T
sorted_indexes = np.argsort(
log_component_probs)[::-1][:top_n]
ordered_word_probs = [(w2v_model.index2word[idx],
log_component_probs[idx])
for idx in sorted_indexes]
print '---'
print "Topic {0}".format(col + 1)
print "Total prob:" + str(sum(log_component_probs))
