diff --git a/skais/learn/hmm_gmm_classifier.py b/skais/learn/hmm_gmm_classifier.py
index 953041d1d5c890b2244ae95233e27680b8c49f55..95e4a5146b5088c1980f3157fffa533358949926 100644
--- a/skais/learn/hmm_gmm_classifier.py
+++ b/skais/learn/hmm_gmm_classifier.py
@@ -30,8 +30,98 @@ def split_trajectories(feature_seq, label_seq, n_classes):
return result, sequence_list
+
+def get_sequences(x, y, n_classes):
+ sequences = {i: [] for i in range(n_classes)}
+ for feature_seq, label_seq in zip(x, y):
+ split_seq, _ = split_trajectories(feature_seq, label_seq, n_classes)
+
+ for key in sequences.keys():
+ sequences[key] += split_seq[key]
+ return sequences
+
+
+def fit_hmmms(self, sequences):
+ for i, seqs in sequences.items():
+ self.hmms[i].n_features = self.n_features
+ if sum([np.array(s).size for s in seqs]) > sum(self.hmms[i]._get_n_fit_scalars_per_param().values()):
+ self.hmms[i].fit(np.concatenate(seqs), list(map(len, seqs)))
+ for j, value in enumerate(self.hmms[i].transmat_.sum(axis=1)):
+ if value == 0:
+ self.hmms[i].transmat_[j][j] = 1.0
+ self.hmms[i].covars_[j] = make_spd_matrix(self.hmms[i].n_features)
+ else:
+ self.hmms[i] = None
+
+
+def get_predictions(x, y, hmms, predictor, n_classes):
+ predict = []
+ for feature_seq, label_seq in zip(x, y):
+ _, sequences_list = split_trajectories(feature_seq, label_seq, n_classes)
+ pred = np.array([])
+ for label, seq in sequences_list:
+ if hmms[label] is not None:
+ _, state_sequence = hmms[label].decode(np.array(seq), [len(seq)])
+ pred = np.append(pred, [predictor[label][i] for i in state_sequence])
+ if len(pred) != 0:
+ predict.append(pred)
+ return predict
+
+
+def get_new_hmm_values(nb_states, predict):
+ start = np.zeros(sum(nb_states))
+ T_mat = np.zeros((sum(nb_states), sum(nb_states)))
+ prior = -1
+ count = np.zeros(sum(nb_states))
+ for pred in predict:
+ start[int(pred[0])] += 1
+
+ for p in pred:
+ if prior != -1:
+ T_mat[prior][int(p)] += 1
+ count[prior] += 1
+ prior = int(p)
+
+ for i in range(sum(nb_states)):
+ for j in range(sum(nb_states)):
+ if T_mat[i][j] > 0:
+ T_mat[i][j] = T_mat[i][j] / count[i]
+
+ for i, value in enumerate(T_mat.sum(axis=1)):
+ if value == 0:
+ T_mat[i][i] = 1.0
+ return start, T_mat, count
+
+
+def get_means_and_covars(hmms, nb_states, nb_features, degens):
+ means = []
+ covars = []
+ for i, model in enumerate(hmms):
+ if hmms[i] is not None:
+ means.append(model.means_)
+ covars.append(model.covars_)
+ else:
+ means.append(np.zeros((nb_states[i], nb_features)))
+ covars.append(np.stack([make_spd_matrix(nb_features)
+ for _ in range(nb_states[i])], axis=0))
+
+ means = np.concatenate(means)
+ covars = np.concatenate(covars)
+ for n, cv in enumerate(covars):
+ if degens[n] and np.any(linalg.eigvalsh(cv) > 0):
+ covars[n] = np.identity(cv.shape[0])
+ limit = 0
+ while (not np.allclose(cv, cv.T) or np.any(linalg.eigvalsh(cv) <= 0)):
+ covars[n] += np.identity(cv.shape[0]) * 10 ** -15
+ if limit > 100:
+ covars[n] = np.identity(cv.shape[0])
+ break
+
+ return means, covars
+
+
class GMMHMMClassifier:
- def __init__(self, nb_states):
+ def __init__(self, nb_states, max_iter=100,verbose=False):
self.n_features = 0
if type(nb_states) is not list:
@@ -39,12 +129,13 @@ class GMMHMMClassifier:
else:
self.nb_states = np.array(nb_states)
+ self.degen_ = [False for _ in range(sum(self.nb_states))]
self.hmms = []
self.n_classes = len(self.nb_states)
for i, nb_state in enumerate(self.nb_states):
- self.hmms.append(GaussianHMM(n_components=nb_state, covariance_type='full'))
+ self.hmms.append(GaussianHMM(n_components=nb_state, covariance_type='full', verbose=verbose, n_iter=max_iter))
- self.hmm = GaussianHMM(n_components=sum(self.nb_states), covariance_type='full', init_params='', n_iter=100)
+ self.hmm = GaussianHMM(n_components=sum(self.nb_states), covariance_type='full', init_params='', n_iter=max_iter)
self.predict_dictionary = {}
self.predictor = []
@@ -57,79 +148,19 @@ class GMMHMMClassifier:
count += 1
def fit(self, x, y):
- sequences = {i: [] for i in range(self.n_classes)}
self.n_features = x[0].shape[1]
- for feature_seq, label_seq in zip(x, y):
- split_seq, _ = split_trajectories(feature_seq, label_seq, self.n_classes)
-
- for key in sequences.keys():
- sequences[key] += split_seq[key]
-
- for i, seqs in sequences.items():
- self.hmms[i].n_features = self.n_features
- if sum([np.array(s).size for s in seqs]) > sum(self.hmms[i]._get_n_fit_scalars_per_param().values()):
- self.hmms[i].fit(np.concatenate(seqs), list(map(len, seqs)))
- for j, value in enumerate(self.hmms[i].transmat_.sum(axis=1)):
- if value == 0:
- self.hmms[i].transmat_[j][j] = 1.0
- self.hmms[i].covars_[j] = make_spd_matrix(self.hmms[i].n_features)
- else:
- self.hmms[i] = None
-
- predict = []
- for feature_seq, label_seq in zip(x, y):
- _, sequences_list = split_trajectories(feature_seq, label_seq, self.n_classes)
- pred = np.array([])
- for label, seq in sequences_list:
- if self.hmms[label] is not None:
- _, state_sequence = self.hmms[label].decode(np.array(seq), [len(seq)])
- pred = np.append(pred, [self.predictor[label][i] for i in state_sequence])
- if len(pred) != 0:
- predict.append(pred)
-
- start = np.zeros(sum(self.nb_states))
- T_mat = np.zeros((sum(self.nb_states), sum(self.nb_states)))
- prior = -1
- count = np.zeros(sum(self.nb_states))
- for pred in predict:
- start[int(pred[0])] += 1
-
- for p in pred:
- if prior != -1:
- T_mat[prior][int(p)] += 1
- count[prior] += 1
- prior = int(p)
-
- for i in range(sum(self.nb_states)):
- for j in range(sum(self.nb_states)):
- if T_mat[i][j] > 0:
- T_mat[i][j] = T_mat[i][j] / count[i]
+
+ sequences = get_sequences(x, y, self.n_classes)
+ fit_hmmms(self, sequences)
+ predict = get_predictions(x, y, self.hmms, self.predictor, self.n_classes)
+ start, T_mat, count = get_new_hmm_values(self.nb_states, predict)
+
+ self.get_degens(count)
self.hmm.startprob_ = start / sum(start)
self.hmm.transmat_ = T_mat
- for i, value in enumerate(self.hmm.transmat_.sum(axis=1)):
- if value == 0:
- self.hmm.transmat_[i][i] = 1.0
-
- means = []
- covars = []
- for i, model in enumerate(self.hmms):
- if self.hmms[i] is not None:
- means.append(model.means_)
- covars.append(model.covars_)
- else:
- means.append(np.zeros((self.nb_states[i], x[0].shape[1])))
- covars.append(np.stack([make_spd_matrix(x[0].shape[1])
- for _ in range(self.nb_states[i])], axis=0))
-
- means = np.concatenate(means)
- covars = np.concatenate(covars)
- for n, cv in enumerate(covars):
- if count[n] <= 3:
- covars[n] = np.identity(cv.shape[0])
- if not np.allclose(cv, cv.T) or np.any(linalg.eigvalsh(cv) <= 0):
- covars[n] += np.identity(cv.shape[0]) * 10 ** -15
+ means, covars = get_means_and_covars(self.hmms, self.nb_states, self.n_features, self.degen_)
self.hmm.means_ = means
self.hmm.covars_ = covars
@@ -146,6 +177,11 @@ class GMMHMMClassifier:
return self.hmm.predict(X_all, lenghts)
+ def get_degens(self, count):
+ for i, c in enumerate(count):
+ if c < self.n_features:
+ self.degen_[i] = True
+
@jit(nopython=True)
def hmm_probabilities(predict, nb_states):
n_states = nb_states.sum()