class ActiveLearner():
##KNN classifier for BOSS \n
#Probably 1 Nearest Neighbour \n
#distance measure: Cosine similarity
class BOSS_NN_classifier(BaseEstimator, ClassifierMixin):
def __init__(self):
self.X_train = None
self.Y_train = None
## fit classifier, for KNN (= lazy learner) store training data
def fit(self, X, Y):
# Check that X and y have correct shape
X, Y = check_X_y(X, Y)
# Store the classes seen during fit
self.classes_ = unique_labels(Y)
self.X_train = X
self.Y_train = Y
## predict confidences of every class for X
def predict_proba(self, X):
# Check if fit had been called
check_is_fitted(self)
# Input validation
X = check_array(X)
similarity = cosine_similarity(X, self.X_train)
neighborclass = self.Y_train[similarity.argmax(axis=1)]
probas = to_categorical(neighborclass,
num_classes=int(max(self.Y_train)) +
1) #+1 für die null
return probas
## predict class of x_test
def predict(self, x_test):
proba_predictedlabels = self.predict_proba(x_test)
predictedlabels = np.argmax(proba_predictedlabels, axis=1)
return predictedlabels
## Rocket Classifier \n
# probably Random Forest
class ROCKETClassifier(RandomForestClassifier):
def __init__(self):
## Settings for Random Forest
super().__init__(
n_estimators=100, max_depth=10, n_jobs=-1
) # Random Forest #warm_start n_estimators=100, max_depth=10,
def __str__(self):
return super().__str__()
## Query Strategy
def imbalance_certainty_sampling(
self, classifier,
X_pool): ## last class is the "don't care" class and gets cut off
probas = self.classifier.predict_proba(X_pool)
probas = probas[:, 0:-1] #cut off "don't cares"
probas = np.max(probas, axis=1)
query_idx = multi_argmax(probas, n_instances=self.queryBagsize)
return query_idx.reshape(-1), X_pool[query_idx].reshape(
self.queryBagsize, -1)
## init function
def __init__(self, windows_pool, n_classes, algo):
self.n_windows = windows_pool.shape[0] # number fo windows
self.samplesPerWindow = windows_pool.shape[
1] #samples per window is need to accelerate ROCKET for faster FST with numba
#super().__init__(n_classes,class_names='', windowLength = 1, samplesPerWindow = samplesPerWindow, n_windows = n_windows,n_initial=0,n_queries=0,algo= algo,query="certainty",fast_mode = False, auto_annotation = False,query_bagsize = 10,timing_run= False, detailResults= False, singleErrorOutput= False)
## TSC Algorrihms
self.algo = algo
## query strategy
self.query_strategy = self.imbalance_certainty_sampling
## bagsize setting
self.queryBagsize = 10
##algo management and Settings
if self.algo == "ROCKET":
# number of kernels Setting
self.n_kernels = 10000 # 2 features per kernel
self.fs_transform = self.Rocket_transform
self.classifier = self.ROCKETClassifier()
## C code generation with numba
_ = generate_kernels(int(self.samplesPerWindow),
int(self.n_kernels))
zeros = np.zeros([self.n_windows, self.samplesPerWindow],
dtype=float)
_ = apply_kernels(np.zeros_like(zeros)[:, 1:], _)
#self.classifier = LogisticRegression()
if self.algo == "BOSS":
self.fs_transform = self.BOSS_transform
# Boss Settings
self.Boss = BOSS(word_size=2,
n_bins=4,
window_size=12,
sparse=False)
#self.Boss = BOSS(word_size=4, n_bins=2, window_size=10, sparse=False)
self.classifier = self.BOSS_NN_classifier()
#self.classifier = RandomForestClassifier()
else:
self.Boss = None
# execute Feature Space Transformation
self.x_pool = self.fs_transform(windows_pool)
# Give IDs to windows
self.x_pool_ID = np.arange(self.x_pool.shape[0])
## Feature Space Transformations
## BOSS FST
def BOSS_transform(self, windows):
X_boss = self.Boss.fit_transform(windows)
return X_boss
## ROCKET FST
def Rocket_transform(self, windows):
#normalized to have a mean of zero and a standard deviation of one
#windows = (windows - windows.mean(axis =1,keepdims=True)) / (windows.std(axis =1,keepdims = True) + 1e-8)
kernels = generate_kernels(windows.shape[1], self.n_kernels)
## C code generation with numba
# wird normalerweise übersprungnen da C Code bereits im Konstruktor generiert wird
_ = generate_kernels(int(self.samplesPerWindow), int(self.n_kernels))
zeros = np.zeros([self.n_windows, self.samplesPerWindow], dtype=float)
_ = apply_kernels(np.zeros_like(zeros)[:, 1:], _)
# apply c code
features = apply_kernels(windows, kernels)
##Scaling for Logistic Regression
#for feature in features:
# feature = preprocessing.scale(feature)
return features
## Real used functions
## search new interesting samples
def query(self):
# modal query
query_idx, query_inst = self.learner.query(self.x_pool)
## queried samples
self.current_queries = query_inst
## and their position in Feature Space
self.current_idx = query_idx
## window IDs = position in storage of windows
windowsIDs = self.x_pool_ID[query_idx]
proba_predictedlabels = self.learner.estimator.predict_proba(
query_inst)
predictedlabels = np.argmax(proba_predictedlabels, axis=1)
return windowsIDs, predictedlabels
## initial training
def initialTraining(self, window_IDs, labels):
#poolIDs = window_IDs because inital training
x_initial, y_initial = self.x_pool[window_IDs], labels
self.learner = modAL_ActiveLearner(estimator=self.classifier,
query_strategy=self.query_strategy,
X_training=x_initial,
y_training=y_initial)
## remove seen samples from pool
self.x_pool, self.x_pool_ID = np.delete(self.x_pool,
window_IDs,
axis=0), np.delete(
self.x_pool_ID,
window_IDs,
axis=0)
## whole Iteration as described in the software architecture "Realization"
def ActiveLearningIteration(self, new_Labels):
# learn new samples with new labels
self.learner.teach(self.current_queries, new_Labels, only_new=False)
# remove learned samples from pool
self.x_pool, self.x_pool_ID = np.delete(self.x_pool,
self.current_idx,
axis=0), np.delete(
self.x_pool_ID,
self.current_idx,
axis=0)
# query new ones
return self.query()
import pytest
import re
from pyts.classification import SAXVSM, BOSSVS
from pyts.multivariate.classification import MultivariateClassifier
from pyts.transformation import BOSS
n_samples, n_features, n_timestamps, n_classes = 40, 3, 30, 2
rng = np.random.RandomState(42)
X = rng.randn(n_samples, n_features, n_timestamps)
y = rng.randint(n_classes, size=n_samples)
@pytest.mark.parametrize(
'params, error, err_msg',
[({
'estimator': [SAXVSM(), SAXVSM(), BOSS()]
}, ValueError, "Estimator 2 must be a classifier."),
({
'estimator': [SAXVSM()]
}, ValueError, "If 'estimator' is a list, its length must be equal to "
"the number of features (1 != 3)"),
({
'estimator': None
}, TypeError, "'estimator' must be a classifier that inherits from "
"sklearn.base.BaseEstimator or a list thereof.")])
def test_parameter_check(params, error, err_msg):
"""Test parameter validation."""
clf = MultivariateClassifier(**params)
with pytest.raises(error, match=re.escape(err_msg)):
clf.fit(X, y)
:class:`pyts.transformation.BOSS`.
"""
import numpy as np
import matplotlib.pyplot as plt
from pyts.transformation import BOSS
# Parameters
n_samples, n_timestamps = 100, 144
# Toy dataset
rng = np.random.RandomState(41)
X = rng.randn(n_samples, n_timestamps)
# BOSS transformation
boss = BOSS(word_size=2, n_bins=4, window_size=12)
X_boss = boss.fit_transform(X).toarray()
# Visualize the transformation for the first time series
plt.figure(figsize=(12, 8))
vocabulary_length = len(boss.vocabulary_)
width = 0.3
plt.bar(np.arange(vocabulary_length) - width / 2,
X_boss[0],
width=width,
label='First time series')
plt.bar(np.arange(vocabulary_length) + width / 2,
X_boss[1],
width=width,
label='Second time series')
plt.xticks(np.arange(vocabulary_length),
It is implemented as :class:`pyts.transformation.BOSS`.
"""
# Author: Johann Faouzi <*****@*****.**>
# License: BSD-3-Clause
import numpy as np
import matplotlib.pyplot as plt
from pyts.datasets import load_gunpoint
from pyts.transformation import BOSS
# Toy dataset
X_train, _, y_train, _ = load_gunpoint(return_X_y=True)
# BOSS transformation
boss = BOSS(word_size=2, n_bins=4, window_size=12, sparse=False)
X_boss = boss.fit_transform(X_train)
# Visualize the transformation for the first time series
plt.figure(figsize=(6, 4))
vocabulary_length = len(boss.vocabulary_)
width = 0.3
plt.bar(np.arange(vocabulary_length) - width / 2,
X_boss[y_train == 1][0],
width=width,
label='First time series in class 1')
plt.bar(np.arange(vocabulary_length) + width / 2,
X_boss[y_train == 2][0],
width=width,
label='First time series in class 2')
plt.xticks(np.arange(vocabulary_length),
# Dynamic Time Warping with a learned warping window
error_dtw_w = 1 - clf_dtw_w.fit(X_train, y_train).score(X_test, y_test)
print('Accuracy DTW_W: ', 1 - error_dtw_w)
print("Error rate with Dynamic Time Warping with a learned warping "
"window: {0:.4f}".format(error_dtw_w))
error_dtw_w_list.append(error_dtw_w)
#BOSS
if dataset_list == ["Adiac"] or dataset_list == ["Herring"]:
window_size = np.repeat(np.arange(60, 110, 20), 3)
norm_mean = np.full(window_size.size, True)
word_size = np.tile(np.arange(10, 16, 2), 3)
boss = [
BOSS(word_size=word_size,
n_bins=4,
norm_mean=norm_mean,
drop_sum=norm_mean,
window_size=window_size)
for (word_size, norm_mean,
window_size) in zip(word_size, norm_mean, window_size)
]
pipeline = [
Pipeline([("boss", boss), ('to_dense', transformer), ("knn", knn)])
for boss in boss
]
voting = VotingClassifier([("pipeline_" + str(i), pipeline)
for i, pipeline in enumerate(pipeline)])
error_boss = 1 - voting.fit(X_train, y_train).score(X_test, y_test)
print('Accuracy BOSS: ', 1 - error_boss)
print("Error rate with BOSS: {0:.4f}".format(error_boss))
error_boss_list.append(error_boss)
import re
from scipy.sparse import csr_matrix
from pyts.classification import SAXVSM
from pyts.image import RecurrencePlot
from pyts.multivariate.transformation import MultivariateTransformer
from pyts.transformation import BOSS
n_samples, n_features, n_timestamps = 40, 3, 30
rng = np.random.RandomState(42)
X = rng.randn(n_samples, n_features, n_timestamps)
@pytest.mark.parametrize(
'params, error, err_msg',
[({
'estimator': [BOSS(), RecurrencePlot(),
SAXVSM()]
}, ValueError, "Estimator 2 must be a transformer."),
({
'estimator': [BOSS()]
}, ValueError, "If 'estimator' is a list, its length must be equal to "
"the number of features (1 != 3)"),
({
'estimator': None
}, TypeError, "'estimator' must be a transformer that inherits from "
"sklearn.base.BaseEstimator or a list thereof.")])
def test_parameter_check(params, error, err_msg):
"""Test parameter validation."""
transformer = MultivariateTransformer(**params)
with pytest.raises(error, match=re.escape(err_msg)):
transformer.fit_transform(X)