def generate_cluster_centers(num_clusters,
hidden_state_dim,
input_dim,
cluster_center_dist_lower_bound,
diagonalizable=True):
"""Generates cluster center eigenvalues with distance requirement.
The generated eigenvalues are drawn uniformly from (-1, 1) until the
pairwise distance between cluster center eigenvalues >= lower bound.
"""
min_cluster_center_dist = -1.
while min_cluster_center_dist < cluster_center_dist_lower_bound:
cluster_centers = []
for _ in xrange(num_clusters):
c = lds.generate_linear_dynamical_system(
hidden_state_dim, input_dim, diagonalizable=diagonalizable)
cluster_centers.append(c)
min_cluster_center_dist = np.inf
for s1 in xrange(num_clusters):
for s2 in xrange(s1 + 1, num_clusters):
d = lds.eig_dist(cluster_centers[s1], cluster_centers[s2])
if d < min_cluster_center_dist:
min_cluster_center_dist = d
logging.info('generated min_cluster_center_dist %.2f',
min_cluster_center_dist)
return cluster_centers
def generate_lds_clusters(cluster_centers,
num_systems,
cluster_radius,
diagonalizable=True):
"""Generates clusters of linear dynamical systems.
Args:
cluster_centers: A list of LinearDynamicalSystem instances.
num_systems: Total number of systems in all clusters.
cluster_radius: Desired mean distance from the centers.
Returns:
- A list of LinearDynamicalSystem of size num_systems.
- A list of true cluster ids.
"""
num_clusters = len(cluster_centers)
cluster_id = np.random.randint(0, num_clusters, num_systems)
hidden_state_dim = cluster_centers[0].hidden_state_dim
for c in cluster_centers:
if c.hidden_state_dim != hidden_state_dim:
raise ValueError('Hidden state dimension mismatch.')
generated_systems = []
dist_to_center = np.zeros(num_systems)
for i in xrange(num_systems):
c = cluster_centers[cluster_id[i]]
if diagonalizable:
eigvalues_new = c.get_spectrum() + cluster_radius / np.sqrt(
hidden_state_dim) * np.random.randn(hidden_state_dim)
generated_systems.append(
lds.generate_linear_dynamical_system(hidden_state_dim,
eigvalues=eigvalues_new))
else:
transition_matrix = c.transition_matrix + cluster_radius / np.sqrt(
hidden_state_dim) * np.random.randn(hidden_state_dim,
hidden_state_dim)
generated_systems.append(
lds.LinearDynamicalSystem(
transition_matrix,
np.random.randn(c.input_matrix.shape[0],
c.input_matrix.shape[1]),
np.random.randn(c.output_matrix.shape[0],
c.output_matrix.shape[1])))
dist_to_center[i] = lds.eig_dist(c, generated_systems[-1])
# For logging purpose.
dist_bw_centers = np.zeros((num_clusters, num_clusters))
for i in xrange(num_clusters):
for j in xrange(num_clusters):
dist_bw_centers[i, j] = lds.eig_dist(cluster_centers[i],
cluster_centers[j])
logging.info('Distances between cluster centers:\n%s',
str(dist_bw_centers))
logging.info('Average distance from cluster centers: %.3f',
np.average(dist_to_center))
for i in xrange(num_clusters):
logging.info('Eigenvalues of cluster center %d: %s', i,
str(cluster_centers[i].get_spectrum()))
return generated_systems, cluster_id
def get_results_seq_len(given_true_eig,
hidden_dim,
input_dim,
min_seq_len,
max_seq_len,
num_sampled_seq_len,
num_repeat,
input_mean,
input_stddev,
output_noise_stddev,
init_state_mean=0.0,
init_state_stddev=0.0,
generate_diagonalizable_only=False,
random_seed=0):
"""Get results for varying sequence lengths.
Args:
given_true_eig: Ground truth of eigenvalues. If None, generate random
eigenvalues from uniform [-1,1] in each repeat of experiment.
hidden_dim: Assumed hidden dim. If 0, use true hidden dim.
input_dim: The input dim.
min_seq_len: Min seq len in experiments.
max_seq_len: Max seq len in experiments.
num_sampled_seq_len: Number of sampled seq len values in between min and max
seq len.
num_repeat: Number of repeated experiments for each seq_len.
input_mean: Scalar or 1D array of length hidden state dim.
input_stddev: Scalar of 1D array of length hidden state dim.
output_noise_stddev: Scalar.
init_state_mean: Scalar or 1D array of length hidden state dim.
init_state_stddev: Scalar of 1D array of length hidden state dim.
generate_diagonalizable_only: Whether to only use diagonalizable LDSs in
simulations.
random_seed: Random seed, integer.
Returns:
A pandas DataFrame with columns `method`, `seq_len`, `t_secs`,
`failed_ratio`, and `l2_r_error`.
The same method and seq_len will appear in num_repeat many rows.
"""
np.random.seed(random_seed)
progress_bar = tqdm.tqdm(total=num_repeat * num_sampled_seq_len)
gen = lds.SequenceGenerator(input_mean=input_mean,
input_stddev=input_stddev,
output_noise_stddev=output_noise_stddev,
init_state_mean=init_state_mean,
init_state_stddev=init_state_stddev)
# seq_len_vals = np.linspace(min_seq_len, max_seq_len, num_sampled_seq_len)
# seq_len_vals = [int(round(x)) for x in seq_len_vals]
min_inv_sqrt_seq_len = 1. / np.sqrt(max_seq_len)
max_inv_sqrt_seq_len = 1. / np.sqrt(min_seq_len)
inv_sqrt_seq_len_vals = np.linspace(min_inv_sqrt_seq_len,
max_inv_sqrt_seq_len,
num_sampled_seq_len)
seq_len_vals = [int(round(1. / (x * x))) for x in inv_sqrt_seq_len_vals]
learning_fns = create_learning_fns(hidden_dim)
metric_dict = {
k: []
for k in [
'method', 'seq_len', 't_secs', 'l2_a_error', 'l2_r_error',
'failed_convg'
]
}
for _ in xrange(num_repeat):
if given_true_eig is not None:
ground_truth = lds.generate_linear_dynamical_system(
hidden_dim, input_dim, eigvalues=given_true_eig)
else:
ground_truth = lds.generate_linear_dynamical_system(
hidden_dim,
input_dim,
diagonalizable=generate_diagonalizable_only)
true_eig = ground_truth.get_spectrum()
for seq_len in seq_len_vals:
seq = gen.generate_seq(ground_truth, seq_len=seq_len)
for k, fn in learning_fns.iteritems():
start_t = timeit.default_timer()
with warnings.catch_warnings(record=True) as caught:
warnings.filterwarnings(
'always', category=sm_exceptions.ConvergenceWarning)
if FLAGS.hide_inputs:
eig_pred = fn(seq.outputs, None)
else:
eig_pred = fn(seq.outputs, seq.inputs)
t_elapsed = timeit.default_timer() - start_t
metric_dict['seq_len'].append(seq_len)
metric_dict['method'].append(k)
metric_dict['t_secs'].append(t_elapsed)
metric_dict['l2_a_error'].append(
np.linalg.norm(true_eig - eig_pred))
metric_dict['l2_r_error'].append(
np.linalg.norm(true_eig - eig_pred) /
np.linalg.norm(true_eig))
metric_dict['failed_convg'].append(False)
for w in caught:
if w.category in [
RuntimeWarning, sm_exceptions.ConvergenceWarning,
sm_exceptions.HessianInversionWarning
]:
metric_dict['failed_convg'][-1] = True
else:
warnings.warn(w.message, w.category)
progress_bar.update(1)
progress_bar.close()
return pd.DataFrame(data=metric_dict)
# See the License for the specific language governing permissions and
# limitations under the License.
"""Script to plot correlation between distances."""
import lds
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
sns.set(style='whitegrid')
num_pairs = 100
plt.figure(figsize=(8, 4))
plt.subplot(1, 2, 1)
hidden_state_dim = 2
lds_pairs = [(lds.generate_linear_dynamical_system(hidden_state_dim),
lds.generate_linear_dynamical_system(hidden_state_dim))
for i in xrange(num_pairs)]
lds_distances = [
lds.eig_dist(system1, system2) for (system1, system2) in lds_pairs
]
expected_ar_distances = [
np.linalg.norm(system1.get_expected_arparams() -
system2.get_expected_arparams())
for (system1, system2) in lds_pairs
]
print(np.corrcoef(lds_distances, expected_ar_distances)[0, 1])
ax = sns.regplot(x=lds_distances, y=expected_ar_distances)
ax.set(xlabel='l-2 distance b/w eigenvalues',
ylabel='l-2 distance b/w '
'corresponding AR params',
def get_results(cluster_center_eigvalues,
cluster_center_dist_lower_bound,
hidden_state_dim,
input_dim,
guessed_hidden_dim,
num_clusters,
guessed_num_clusters,
min_seq_len,
max_seq_len,
num_sampled_seq_len,
num_repeat,
num_systems,
cluster_radius,
input_mean,
input_stddev,
output_noise_stddev,
init_state_mean=0.0,
init_state_stddev=0.0,
generate_diagonalizable_only=False,
random_seed=0,
results_path=None):
"""Get results for varying sequence lengths.
Args:
cluster_center_eigvalues: List of lists of eigenvalues for each cluster.
E.g. [[0.9,0.1], [0.5,0.1], [0.2,0.2], or None. If None, eigenvalues will
be generated from uniform(-1,1) with respect to
cluster_center_dist_lower_bound.
cluster_center_dist_lower_bound: Desired distance lower bound between
clusters. When generating cluster centers, try repeatedly until distance
is greater than cluster_center_dist_lower_bound.
hidden_state_dim: True hidden state dim.
input_dim: The input dim.
guessed_hidden_dim: Assumed hidden dim. If 0, use true hidden dim.
num_clusters: True number of clusters.
guessed_num_clusters: Desired number of clusters. If 0, use true number.
min_seq_len: Min seq len in experiments.
max_seq_len: Max seq len in experiments.
num_sampled_seq_len: Number of sampled seq len values in between min and max
seq len.
num_repeat: Number of repeated experiments for each seq_len.
num_systems: Number of dynamical system in each clustering experiments.
cluster_radius: Expected distance of generated systems from cluster centers.
input_mean: Scalar or 1D array of length hidden state dim.
input_stddev: Scalar of 1D array of length hidden state dim.
output_noise_stddev: Scalar.
init_state_mean: Scalar or 1D array of length hidden state dim.
init_state_stddev: Scalar of 1D array of length hidden state dim.
random_seed: Random seed, integer.
Returns:
A pandas DataFrame with columns `method`, `seq_len`, `t_secs`,
`failed_ratio`, and columns for clustering metrics such as `adj_mutual_info`
and `v_measure`. The same method and seq_len will appear in num_repeat many
rows.
"""
if cluster_center_eigvalues is not None:
if len(cluster_center_eigvalues) <= 1:
raise ValueError('Need at least two cluster centers.')
cluster_center_eigvalues = np.array(cluster_center_eigvalues)
if cluster_center_eigvalues.shape != (num_clusters, hidden_state_dim):
raise ValueError(
'Cluter center eig has shape %s, expected (%d, %d).' %
(str(cluster_center_eigvalues.shape), num_clusters,
hidden_state_dim))
np.random.seed(random_seed)
progress_bar = tqdm.tqdm(total=num_repeat * num_sampled_seq_len)
# Generator for output sequences.
gen = lds.SequenceGenerator(input_mean=input_mean,
input_stddev=input_stddev,
output_noise_stddev=output_noise_stddev,
init_state_mean=init_state_mean,
init_state_stddev=init_state_stddev)
seq_len_vals = np.linspace(min_seq_len, max_seq_len, num_sampled_seq_len)
seq_len_vals = [int(round(x)) for x in seq_len_vals]
if guessed_hidden_dim == 0:
guessed_hidden_dim = hidden_state_dim
if guessed_num_clusters == 0:
guessed_num_clusters = num_clusters
results_dfs = []
for i in xrange(num_repeat):
logging.info('---Starting experiments in repeat run #%d---', i)
if cluster_center_eigvalues is not None:
cluster_centers = []
for eig_val in cluster_center_eigvalues:
c = lds.generate_linear_dynamical_system(hidden_state_dim,
input_dim,
eigvalues=eig_val)
cluster_centers.append(c)
else:
cluster_centers = clustering.generate_cluster_centers(
num_clusters,
hidden_state_dim,
input_dim,
cluster_center_dist_lower_bound,
diagonalizable=generate_diagonalizable_only)
true_systems, true_cluster_ids = clustering.generate_lds_clusters(
cluster_centers,
num_systems,
cluster_radius,
diagonalizable=generate_diagonalizable_only)
for seq_len in seq_len_vals:
logging.info('Running experiment with seq_len = %d.', seq_len)
seqs = [gen.generate_seq(s, seq_len=seq_len) for s in true_systems]
# Create transform_fns.
model_fns = create_model_fns(guessed_hidden_dim)
pca = sklearn.decomposition.PCA(
n_components=guessed_hidden_dim).fit(
np.stack([s.outputs.flatten() for s in seqs], axis=0))
model_fns['PCA'] = _create_pca_model_fn(pca)
transform_fns = collections.OrderedDict()
for k in model_fns:
transform_fns[k] = _compose_model_fn(model_fns[k])
# Get clustering results.
results_df = clustering.get_results(
seqs,
guessed_num_clusters,
true_cluster_ids,
true_systems,
transform_fns,
FLAGS.include_tslearn,
include_slow_methods=FLAGS.include_tslearn_slow)
results_df['seq_len'] = seq_len
results_df['n_guessed_clusters'] = guessed_num_clusters
results_df['n_true_clusters'] = num_clusters
results_df['true_hidden_dim'] = hidden_state_dim
results_df['guessed_hidden_dim'] = guessed_hidden_dim
results_dfs.append(results_df)
logging.info('Results:\n%s', str(results_df))
plot_filepath = os.path.join(
FLAGS.output_dir,
'cluster_visualization_run_%d_seq_len_%d.png' % (i, seq_len))
if FLAGS.plot_clusters:
clustering.visualize_clusters(seqs, true_systems,
true_cluster_ids, transform_fns,
plot_filepath)
progress_bar.update(1)
if results_path:
with open(results_path, 'w+') as f:
pd.concat(results_dfs).to_csv(f, index=False)
progress_bar.close()
return pd.concat(results_dfs)
