def fit(X, y, max_depth=None, min_samples_split=2, min_len=10, max_len=10):
print('New node')
if (max_depth is None or max_depth > 0) and len(X) > min_samples_split and len(np.unique(y)) > 1:
# TODO: pass the distance along with this shapelet so we don't need to recalculate this!
shapelet = fast_shapelet_discovery(X, y,min_len=min_len,max_len=max_len)
distance = check_candidate(X, y, shapelet)[1]
node = ShapeletTree(right=None, left=None, shapelet=shapelet, distance=distance,
class_probabilities=Counter(y))
X_left, y_left, X_right, y_right = [], [], [], []
for ts, label in zip(X, y):
if subsequence_dist(ts, shapelet)[0] <= distance:
X_left.append(ts)
y_left.append(label)
else:
X_right.append(ts)
y_right.append(label)
new_depth = None if max_depth is None else max_depth - 1
node.left = fit(X_left, y_left, max_depth=new_depth, min_samples_split=min_samples_split,
min_len=min_len, max_len=max_len)
node.right = fit(X_right, y_right, max_depth=new_depth, min_samples_split=min_samples_split,
min_len=min_len, max_len=max_len)
return node
else:
return ShapeletTree(right=None, left=None, shapelet=None, distance=None,
class_probabilities=Counter(y))
def increment_class_probs(self, ts, label):
if label not in self.class_probabilities:
self.class_probabilities[label] = 1
else:
self.class_probabilities[label] += 1
if self.distance is not None:
dist, idx = subsequence_dist(ts, self.shapelet)
if dist <= self.distance:
self.left.increment_class_probs(ts, label)
else:
self.right.increment_class_probs(ts, label)
def evaluate(self, time_serie, proba=True):
if self.distance is None:
if proba:
return self.class_probabilities
else:
return max(self.class_probabilities.items(),
key=operator.itemgetter(1))[0]
else:
dist, idx = subsequence_dist(time_serie, self.shapelet)
if dist <= self.distance:
return self.left.evaluate(time_serie, proba=proba)
else:
return self.right.evaluate(time_serie, proba=proba)
def check_candidate(timeseries,
labels,
shapelet,
min_prune_length=20,
best_ig=None):
distances = []
# cntr = Counter(labels)
for time_serie, label in zip(timeseries, labels):
d, idx = util.subsequence_dist(time_serie, shapelet)
distances.append((d, label))
# max_ig = None
# if best_ig is not None:
# cntr[label] -= 1
# if len(distances) > min_prune_length:
# max_ig = entropy_pre_prune(cntr, distances)
# if max_ig is not None and max_ig <= best_ig:
# return 0, 0
return find_best_split_point(sorted(distances, key=lambda x: x[0]))
def recalculate_distances(self, timeseries, labels):
if self.distance is not None:
ig, dist = check_candidate(timeseries, labels, self.shapelet)
print(dist, self.distance)
self.distance = dist
ts_left, labels_left = [], []
ts_right, labels_right = [], []
for (ts, label) in zip(timeseries, labels):
dist, idx = subsequence_dist(ts, self.shapelet)
if dist < self.distance:
ts_left.append(ts)
labels_left.append(label)
else:
ts_right.append(ts)
labels_right.append(label)
print(labels, 'are split into', labels_left, 'and', labels_right)
self.left.recalculate_distances(ts_left, labels_left)
self.right.recalculate_distances(ts_right, labels_right)
else:
print('leaf:', labels)
# assert np.array_equal(m_uv, m_uv_old)
print(labels)
print(timeseries)
if __name__ == "__main__":
print('Fitting tree')
tree = extract_shapelet(timeseries, labels)
print(tree.shapelet)
print(tree.distance)
distances = []
for ts, label in zip(timeseries, labels):
d, idx = subsequence_dist(ts, tree.shapelet)
distances.append((d, label))
print([x for x in sorted(distances, key=lambda x: x[0])])
distances = []
for ts, label in zip(timeseries, labels):
stats = calculate_stats(tree.shapelet, ts)
d = sdist_new(tree.shapelet, ts, 0, stats)
distances.append((d, label))
print([x for x in sorted(distances, key=lambda x: x[0])])
distances = []
stats[tuple(ts2)] = util.calculate_stats(ts, ts2)
sdist_new_overhead.append(time.time() - start_time)
for l in range(1, ts_length + 1):
for start in range(len(ts) - l): # Possible start positions
new_dists = []
old_dists = []
for k, (ts2, label2) in enumerate(zip(timeseries, labels)):
start_time = time.time()
dist_new = util.sdist_new(ts[start:start + l], ts2,
start, stats[tuple(ts2)])
sdist_new_times.append(time.time() - start_time)
new_dists.append((k, dist_new))
start_time = time.time()
dist_old, idx_old = util.subsequence_dist(
ts2, ts[start:start + l])
sdist_old_times.append(time.time() - start_time)
old_dists.append((k, dist_old))
new_dists = sorted(new_dists, key=lambda x: x[1])
old_dists = sorted(old_dists, key=lambda x: x[1])
print(new_dists)
print(old_dists)
np.testing.assert_equal([x[0] for x in new_dists][0],
[x[0] for x in old_dists][0])
print('New distance calculation took:',
np.sum(sdist_new_times) + np.sum(sdist_new_overhead))
print('New distance (overhead):', np.sum(sdist_new_overhead))
print('Old distance calculation took:', np.sum(sdist_old_times))