ethomap 0.1.1

Generate behavioral embeddings of postural time series.

Ethomap

Lightweight library with basic modules for generating behavioral embeddings from dynamic time warping of postural time series. Also includes modules to decompose behavioral sequences into transition modes.

Based on Mearns et al. (2020).

Installation

pip install ethomap

Examples

Dynamic time warping

from ethomap import DynamicTimeWarping  # from dtw module
import numpy as np

# Create some 3D time series
x = np.random.rand(200, 3)
y = np.random.rand(200, 3)

# Align time series y to x
dtw = DynamicTimeWarping(x)
dist = dtw.align(y)
print("DTW distance between x and y:", dist)

Behavioral embedding

from ethomap import pdist_dtw, affinity_propagation, IsomapPrecomputed
import numpy as np
import time
from scipy.spatial.distance import squareform

# Create some 3D time series to align pairwise
xs = np.random.rand(1000, 50, 3)

# Compute pairwise distances between time series (pdist_dtw from ethomap.distance)
t0 = time.time()
distances = pdist_dtw(xs, parallel_processing=True, n_processors=4, bw=0.1, fs=100)
distance_matrix = squareform(distances)  # make square
t1 = time.time()
print("Time taken:", t1 - t0, "seconds")  # ~30 seconds for data of this size
print(distance_matrix.shape)  # (1000, 1000)

# Create behavioral space from all data (IsomapPrecomputed from ethomap.embed)
isomap = IsomapPrecomputed(n_neighbors=5, n_components=2)
embedding = isomap.fit_transform(distance_matrix)
print(embedding.shape)  # (1000, 2)

# Find exemplars by clustering distance matrix (affinity_propagation from ethomap.cluster)
ap = affinity_propagation(distance_matrix, preference="median")
cluster_labels = ap.labels_
exemplar_indices = ap.cluster_centers_indices_

# Select exemplars from original data
exemplars = xs[exemplar_indices]
exemplar_distances = distance_matrix[exemplar_indices, :][:, exemplar_indices]
print(exemplars.shape)  # (n_exemplars, 50, 3)
print(exemplar_distances.shape)  # (n_exemplars, n_exemplars)

# Create behavioral space from exemplars only
embedding_exemplars = isomap.fit_transform(exemplar_distances)
print(embedding_exemplars.shape)  # (n_exemplars, 2)