def test_weighted_var():
es = EventSegment(2)
D = np.zeros((8, 4))
for t in range(4):
D[t, :] = (1/np.sqrt(4/3)) * np.array([-1, -1, 1, 1])
for t in range(4, 8):
D[t, :] = (1 / np.sqrt(4 / 3)) * np.array([1, 1, -1, -1])
mean_pat = D[[0, 4], :].T
weights = np.zeros((8, 2))
weights[:, 0] = [1, 1, 1, 1, 0, 0, 0, 0]
weights[:, 1] = [0, 0, 0, 0, 1, 1, 1, 1]
assert np.array_equal(
es.calc_weighted_event_var(D, weights, mean_pat), [0, 0]),\
"Failed to compute variance with 0/1 weights"
weights[:, 0] = [1, 1, 1, 1, 0.5, 0.5, 0.5, 0.5]
weights[:, 1] = [0.5, 0.5, 0.5, 0.5, 1, 1, 1, 1]
true_var = (4 * 0.5 * 12)/(6 - 5/6) * np.ones(2) / 4
assert np.allclose(
es.calc_weighted_event_var(D, weights, mean_pat), true_var),\
"Failed to compute variance with fractional weights"
def test_weighted_var():
es = EventSegment(2)
D = np.zeros((8, 4))
for t in range(4):
D[t, :] = (1 / np.sqrt(4 / 3)) * np.array([-1, -1, 1, 1])
for t in range(4, 8):
D[t, :] = (1 / np.sqrt(4 / 3)) * np.array([1, 1, -1, -1])
mean_pat = D[[0, 4], :].T
weights = np.zeros((8, 2))
weights[:, 0] = [1, 1, 1, 1, 0, 0, 0, 0]
weights[:, 1] = [0, 0, 0, 0, 1, 1, 1, 1]
assert np.array_equal(
es.calc_weighted_event_var(D, weights, mean_pat), [0, 0]),\
"Failed to compute variance with 0/1 weights"
weights[:, 0] = [1, 1, 1, 1, 0.5, 0.5, 0.5, 0.5]
weights[:, 1] = [0.5, 0.5, 0.5, 0.5, 1, 1, 1, 1]
true_var = (4 * 0.5 * 12) / (6 - 5 / 6) * np.ones(2) / 4
assert np.allclose(
es.calc_weighted_event_var(D, weights, mean_pat), true_var),\
"Failed to compute variance with fractional weights"
#### 3.2 Define events from movie, find in recall Alternatively, rather than finding events simultaneously in the movie and recall, we can first identify events using the movie data only, and then go looking for this fixed set of events in the recall. In this case we can use other methods, such as GSBS, to identify the events in the movie, and then use the HMM to transfer these events to the recall. Below we do this for both the 40-event and 109-event segmentations from GSBS. Note that we need to specify a recall event variance for the HMM when applying it in this way - this value controls how confident the model is about whether it knows which event is being recalled given a spatial pattern. This parameter could be cross-validated relative to some other measure of performance, but here we just set the recall variance to be the same as the movie variance for each event. # Use 40 GSBS events to set the HMM event patterns n_events = 40 event_pat = GSBS_states.get_state_patterns(n_events).T state_labels = GSBS_states.get_states(n_events) movie_events40 = np.zeros((nTRs, np.max(state_labels))) movie_events40[np.arange(nTRs), state_labels-1] = 1 HMM = EventSegment(n_events) HMM.set_event_patterns(event_pat) ev_var = HMM.calc_weighted_event_var(movie_group, movie_events40, HMM.event_pat_) recall_events40 = HMM.find_events(recall, var = ev_var)[0] # Use 109 GSBS events to set the HMM event patterns n_events = GSBS_states.nstates event_pat = GSBS_states.state_patterns.T state_labels = GSBS_states.get_states() movie_events109 = np.zeros((nTRs, np.max(state_labels))) movie_events109[np.arange(nTRs), state_labels-1] = 1 HMM = EventSegment(n_events) HMM.set_event_patterns(event_pat) ev_var = HMM.calc_weighted_event_var(movie_group, movie_events109, HMM.event_pat_) recall_events109 = HMM.find_events(recall, var = ev_var)[0]