def compute_SVD(T, S):
T_USVs, T_USVa = SVD(T)
S_USVs, S_USVa = SVD(S)
Ts = np.diag(T_USVs[1])[:10]
Ta = np.diag(T_USVa[1])[:20:2]
Ss = np.diag(S_USVs[1])[:10]
Sa = np.diag(S_USVa[1])[:20:2]
return np.array([Ts, Ta, Ss, Sa])
def compare_transition_modes(m1, m2, exact=False, n_permutations=1000):
"""Compare the transitions matrices between two groups of animals.
Parameters
----------
m1, m2: np.ndarray, shape (n_animals, n_states, n_states)
Transition matrices for the two groups of animals
exact : bool, default False
Whether to compute every possible permutation
n_permutations
Number of permutations to compute if exact is False
Returns
-------
dot_products : np.ndarray, shape (N, 3)
N x 3 matrix containing dot products between first two symmetric and first antisymmetric transition modes across
N permutations of animals
"""
# Number of animals in test group
n_test = len(m1)
# Concatenate transition matrices
all_transitions = np.concatenate([m1, m2], axis=0)
idxs = np.arange(len(all_transitions))
# Output
dot_products = []
# Iterate through different permutations of animals
i = 0
for group1, group2 in permute(idxs,
n_test,
exact=exact,
n_permutations=n_permutations):
if i % 10 == 0:
print i
# Sum transition matrices
T1 = all_transitions[group1].sum(axis=0)
T2 = all_transitions[group2].sum(axis=0)
# Compute the SVD
USVs1, USVa1 = SVD(T1)
USVs2, USVa2 = SVD(T2)
# Compare first two symmetric transition modes
s1_dot = np.abs(np.dot(USVs1[0, :, 0], USVs2[0, :, 0]))
s2_dot = np.abs(np.dot(USVs1[0, :, 1], USVs2[0, :, 1]))
# Compare first antisymmetric transition mode
a1_dot_a = np.abs(np.dot(USVa1[0, :, 0], USVa2[0, :, 0]))
a1_dot_b = np.abs(np.dot(USVa1[0, :, 0], USVa2[0, :, 1]))
a1_dot = max(a1_dot_a, a1_dot_b)
a2_dot_a = np.abs(np.dot(USVa1[0, :, 1], USVa2[0, :, 0]))
a2_dot_b = np.abs(np.dot(USVa1[0, :, 1], USVa2[0, :, 1]))
a2_dot = max(a2_dot_a, a2_dot_b)
# Append to output
dot_products.append((s1_dot, s2_dot, a1_dot, a2_dot))
i += 1
return np.array(dot_products)
transition_directory = create_folder(experiment.subdirs['analysis'],
'transitions')
for condition, fish_info in experiment.data.groupby('condition'):
print condition
condition_directory = create_folder(transition_directory, condition)
condition_bouts = mapped_bouts[mapped_bouts['ID'].isin(
fish_info['ID'])]
# Compute the transition matrix for each fish
T = fish_transition_matrices(condition_bouts,
state_col='exemplar',
n_states=len(isomap),
shuffle=False)
np.save(os.path.join(condition_directory, 'transition_matrices.npy'),
T)
print T.shape
# Redistribute transitions
W = generate_weights(isomap, bandwidth=40.)
WTW = redistribute_transitions(T, W)
np.save(
os.path.join(condition_directory,
'smoothed_transition_matrices.npy'), WTW)
# Sum transitions over all fish
T_all = T.sum(axis=0)
WTW_all = redistribute_transitions(T_all, W)
np.save(os.path.join(condition_directory, 'T.npy'), T_all)
np.save(os.path.join(condition_directory, 'WTW.npy'), WTW_all)
# SVD
USVs, USVa = SVD(WTW_all)
np.save(os.path.join(condition_directory, 'USVs.npy'), USVs)
np.save(os.path.join(condition_directory, 'USVa.npy'), USVa)
'smoothed_transition_matrices.npy'))
unilateral_matrices = np.load(
os.path.join(transition_directory, 'unilateral',
'smoothed_transition_matrices.npy'))
# timer = Timer()
# timer.start()
# dot_products = compare_transition_modes(unilateral_matrices, control_matrices, exact=False, n_permutations=1000)
# np.save(os.path.join(transition_directory, 'compare_control_unilateral.npy'), dot_products)
# print timer.convert_time(timer.stop())
# print '\n'
T1 = np.sum(control_matrices, axis=0)
T2 = np.sum(unilateral_matrices, axis=0)
USVs1, USVa1 = SVD(T1)
USVs2, USVa2 = SVD(T2)
# Compare first two symmetric transition modes
s1_dot = np.abs(np.dot(USVs1[0, :, 0], USVs2[0, :, 0]))
s2_dot = np.abs(np.dot(USVs1[0, :, 1], USVs2[0, :, 1]))
# Compare first antisymmetric transition mode
a1_dot_a = np.abs(np.dot(USVa1[0, :, 0], USVa2[0, :, 0]))
a1_dot_b = np.abs(np.dot(USVa1[0, :, 0], USVa2[0, :, 1]))
a1_dot = max(a1_dot_a, a1_dot_b)
a2_dot_a = np.abs(np.dot(USVa1[0, :, 1], USVa2[0, :, 0]))
a2_dot_b = np.abs(np.dot(USVa1[0, :, 1], USVa2[0, :, 1]))
a2_dot = max(a2_dot_a, a2_dot_b)
print s1_dot, s2_dot
print a1_dot, a2_dot
from datasets.main_dataset import experiment
from behaviour_analysis.analysis.transitions import SVD
import numpy as np
import os
if __name__ == "__main__":
transition_directory = os.path.join(experiment.subdirs['analysis'], 'transitions')
WTW = np.load(os.path.join(transition_directory, 'WTW.npy'))
USVs, USVa = SVD(WTW)
np.save(os.path.join(transition_directory, 'USVs.npy'), USVs)
np.save(os.path.join(transition_directory, 'USVa.npy'), USVa)