def fit(self, X, Y):
Xs_train = []
Ys_train = []
for i in range(len(X)):
Xs, Ys = utils.segment_data(X[i], Y[i])
if self.normalize:
Xs = [normalize(x) for x in Xs]
Xs_train += Xs
Ys_train += Ys
self.Xs_train = Xs_train
self.Ys_train = Ys_train
self.n_train = len(Xs_train)
self.n_classes = len(np.unique(Ys_train))
def fit(self, X, Y):
Xs_train = []
Ys_train = []
for i in range(len(X)):
Xs, Ys = utils.segment_data(X[i], Y[i])
if self.normalize:
Xs = [normalize(x) for x in Xs]
Xs_train += Xs
Ys_train += Ys
self.Xs_train = Xs_train
self.Ys_train = Ys_train
self.n_train = len(Xs_train)
self.n_classes = len(np.unique(Ys_train))
def predict(self, X, Y, return_prob=False):
if type(X) is list:
out = []
for i in range(len(Xi)):
out += [self.predict(model, X[i], Y[i], return_prob)]
return out
Xs, Ys = utils.segment_data(X, Y)
n_segs = len(Xs)
if self.normalize:
Xs = [normalize(x) for x in Xs]
Ps = np.zeros(n_segs, np.int)
Ps_prob = np.zeros([self.n_classes, n_segs], np.float)
top_scores = [np.inf] * self.n_neighbors
for k in range(n_segs):
scores = np.zeros(self.n_train, np.float)
worst_best_score = np.inf
for j in range(self.n_train):
scores[j] = DTW(Xs[k][:, ::self.sample_rate],
self.Xs_train[j][:, ::self.sample_rate],
worst_best_score)
if scores[j] < np.min(top_scores):
idx = np.argmin(top_scores)
top_scores[idx] = scores[j]
worst_best_score = np.max(top_scores)
argsorted = scores.argsort()
neighbors = [
self.Ys_train[a] for a in argsorted[:self.n_neighbors]
]
neighbors_scores = np.array(
[scores[a] for a in argsorted[:self.n_neighbors]])
Ps_prob[:, k] = np.histogram(neighbors, self.n_classes,
[0, self.n_classes])[0] * 1.
Ps_prob[:, k] /= Ps_prob[:, k].sum()
Ps[k] = scipy.stats.mode(neighbors)[0][0]
if return_prob:
return Ps_prob
else:
return Ps
def predict(self, X, Y, return_prob=False):
if type(X) is list:
out = []
for i in range(len(Xi)):
out += [self.predict(model, X[i], Y[i], return_prob)]
return out
Xs, Ys = utils.segment_data(X, Y)
n_segs = len(Xs)
if self.normalize:
Xs = [normalize(x) for x in Xs]
Ps = np.zeros(n_segs, np.int)
Ps_prob = np.zeros([self.n_classes, n_segs], np.float)
top_scores = [np.inf]*self.n_neighbors
for k in range(n_segs):
scores = np.zeros(self.n_train, np.float)
worst_best_score = np.inf
for j in range(self.n_train):
scores[j] = DTW(Xs[k][:,::self.sample_rate], self.Xs_train[j][:,::self.sample_rate], worst_best_score)
if scores[j] < np.min(top_scores):
idx = np.argmin(top_scores)
top_scores[idx] = scores[j]
worst_best_score = np.max(top_scores)
argsorted = scores.argsort()
neighbors = [self.Ys_train[a] for a in argsorted[:self.n_neighbors]]
neighbors_scores = np.array([scores[a] for a in argsorted[:self.n_neighbors]])
Ps_prob[:,k] = np.histogram(neighbors, self.n_classes, [0, self.n_classes])[0]*1.
Ps_prob[:,k] /= Ps_prob[:,k].sum()
Ps[k] = scipy.stats.mode(neighbors)[0][0]
if return_prob:
return Ps_prob
else:
return Ps
def clf_(p, y):
segs_p, segs_y = utils.segment_data(p[:,None].T, y)
segs = np.array([scipy.stats.mode(s)[0][0] for s in segs_p])
return np.mean(segs == segs_y)*100
