ad_examples/python/timeseries/timeseries_regression.py at master · Wolgo/ad_examples · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
import random
import numpy as np
import numpy.random as rnd
from common.utils import *
from common.data_plotter import *
from common.nn_utils import *

from sklearn import svm
from sklearn.ensemble import RandomForestRegressor
from sklearn.neural_network import MLPRegressor as MLPRegressor_SK

from common.timeseries_datasets import *
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'


"""
To execute:
pythonw -m timeseries.timeseries_regression --dataset=airline --algo=nntf --n_lags=12 --n_epochs=200 --debug --log_file=temp/timeseries/timeseries_regression.log --normalize_trend
pythonw -m timeseries.timeseries_regression --dataset=shampoo --algo=nntf --n_lags=10 --n_epochs=200 --debug --log_file=temp/timeseries/timeseries_regression.log --normalize_trend
pythonw -m timeseries.timeseries_regression --dataset=lynx --algo=nntf --n_lags=20 --n_epochs=200 --debug --log_file=temp/timeseries/timeseries_regression.log
pythonw -m timeseries.timeseries_regression --dataset=aus_beer --algo=nntf --n_lags=12 --n_epochs=200 --debug --log_file=temp/timeseries/timeseries_regression.log
pythonw -m timeseries.timeseries_regression --dataset=us_accident --algo=nntf --n_lags=12 --n_epochs=200 --debug --log_file=temp/timeseries/timeseries_regression.log
pythonw -m timeseries.timeseries_regression --dataset=wolf_sunspot --algo=nnsk --n_lags=50 --n_epochs=200 --debug --log_file=temp/timeseries/timeseries_regression.log

pythonw -m timeseries.timeseries_regression --dataset=fisher_temp --algo=nntf --n_lags=20 --n_epochs=100 --debug --log_file=temp/timeseries/timeseries_regression.log
"""


def find_anomalies_with_regression(data, dataset, n_lags=5, reg_type="svr", n_anoms=10,
                                   normalize_trend=False, batch_size=10, n_epochs=200, tr_frac=0.8):
    """ Finds anomalous time points in time series using standard regression algorithms

    SVR might be fine even if trend is not removed, but when using
    RandomForestRegressor, make sure that the trend is removed.
    Ideally, remove trend and standardize for all timeseries.
    """

    n = data.shape[0]
    n_tr = int(n * tr_frac)

    if normalize_trend:
        # remove trend from training series
        diff_data = difference_series(data)
    else:
        diff_data = data
    diff_tr = diff_data[:n_tr]
    diff_test = diff_data[n_tr:]  # will be used to score predictions

    # normalize to range (-1, 1)
    normalizer = DiffScale()
    scaled_tr = normalizer.fit_transform(diff_tr, normalize_trend=False)
    scaled_test = normalizer.scale(diff_test)  # will be used to score predictions later

    ts = prepare_tseries(scaled_tr, name=dataset)

    if False:
        logger.debug("diff_data:(%s, %s, %s, %s)\n%s\n%s" %
                     (str(data.shape), str(diff_data.shape), str(diff_tr.shape), str(diff_test.shape),
                      str(scaled_tr[:, 0]), str(scaled_test[:, 0])))

    if False:
        # debug only
        logger.debug("samples:(%s)\n%s" % (str(ts.samples.shape), str(ts.samples[:, 0])))
        for x, y in ts.get_batches(n_lags=n_lags, batch_size=-1, single_output_only=True):
            batch = np.hstack([y, np.reshape(x, newshape=(x.shape[0], -1))])
            logger.debug("batch:\n%s" % str(batch))
        return

    x = y = None
    for x_, y_ in ts.get_batches(n_lags, batch_size=-1, single_output_only=True):
        x = np.reshape(x_, newshape=(x_.shape[0], -1))
        y = np.reshape(y_, newshape=(-1,))

    if False:
        logger.debug("Total train instances: %d" % x.shape[0])
        logger.debug("y:\n%s" % str(y))

    if reg_type == "svr":
        mdl = svm.SVR(C=1.0, cache_size=200, coef0=0.0, degree=3, epsilon=0.1, gamma='auto',
                      kernel='linear', max_iter=-1, shrinking=True, tol=0.001, verbose=False)
    elif reg_type == "rfor":
        mdl = RandomForestRegressor(n_estimators=10, criterion="mse", max_depth=None,
                                    min_samples_split=2, min_samples_leaf=1, min_weight_fraction_leaf=0.0,
                                    max_features="auto", max_leaf_nodes=None, min_impurity_decrease=0.0,
                                    min_impurity_split=None, bootstrap=True, oob_score=False, n_jobs=1,
                                    random_state=None, verbose=0, warm_start=False)
    elif reg_type == "nntf":
        # use tensorflow
        mdl = MLPRegressor_TF(x.shape[1], 100, 1, batch_size=batch_size, shuffle=True,
                              n_epochs=n_epochs, l2_penalty=0.01)
    elif reg_type == "nnsk":
        # use scikit-learn
        mdl = MLPRegressor_SK(hidden_layer_sizes=(100, ), activation='relu', solver='adam',
                              alpha=0.0001, batch_size='auto', learning_rate='constant',
                              learning_rate_init=0.001, power_t=0.5, max_iter=n_epochs, shuffle=True,
                              random_state=None, tol=0.0001, verbose=False, warm_start=False,
                              momentum=0.9, nesterovs_momentum=True, early_stopping=False,
                              validation_fraction=0.1, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
    else:
        raise ValueError("unsupported regression type: %s" % reg_type)

    # remember: x, y are only the train series
    mdl.fit(x, y)

    # Predict on test set. Assumes that all previous values at a time point as known
    pred_points = list(scaled_tr[:, 0])
    prev_points = list(scaled_tr[:, 0])
    for i in range(len(scaled_test)):
        yhat = mdl.predict(np.reshape(np.array(prev_points[-n_lags:]), newshape=(1, n_lags)))[0]
        logger.debug("%d yhat: %f" % (i, yhat))
        pred_points.append(yhat)
        prev_points.append(scaled_test[i, 0])  # to predict just one step ahead of known values
        # prev_points.append(yhat)  # to predict far into the future

    logger.debug("pred_points:(%d)\n%s" % (len(pred_points[scaled_tr.shape[0]:]), str(list(pred_points[scaled_tr.shape[0]:]))))

    # now invert all transformations
    pred_points = np.reshape(pred_points, newshape=(-1, ts.y.shape[1]))
    pred_points = normalizer.inverse_transform(pred_points)
    if normalize_trend:
        pred_points = invert_difference_series(pred_points, data[[0], :])

    scores = np.abs(data[n_tr:, 0] - pred_points[n_tr:, 0])

    top_anoms = np.argsort(-scores)[0:n_anoms]
    logger.debug("top scores (%s):\n%s\n%s" % (reg_type, str(top_anoms), str(scores[top_anoms])))

    pdfpath = "temp/timeseries/timeseries_regression_%s_%d_%s.pdf" % (ts.name, n_lags, reg_type)
    dp = DataPlotter(pdfpath=pdfpath, rows=3, cols=1)
    pl = dp.get_next_plot()
    pl.set_xlim([0, data.shape[0]])
    pl.plot(np.arange(0, n), data[:, 0], 'b-', linewidth=0.5)
    # pl.plot(np.arange(0, n), pred_points[:, 0], 'g-', linewidth=0.5)
    pl.plot(np.arange(n_tr, n), pred_points[n_tr:, 0], 'r-', linewidth=0.5)
    for i in top_anoms:
        plt.axvline(i+n_tr, color='g', linewidth=0.5)
    dp.close()


def read_ts():
    samples = pd.read_csv("../datasets/simulated_timeseries/samples_2000.csv",
                          header=None, sep=",", usecols=[1]
                          )
    samples = np.asarray(samples, dtype=np.float32)
    return TSeries(samples[:2000, :], y=None)


if __name__ == "__main__":

    logger = logging.getLogger(__name__)

    args = get_command_args(debug=False,
                            debug_args=["--dataset=airline", "--algo=nntf", "--n_lags=12",
                                        "--n_anoms=10", "--debug", "--plot",
                                        "--log_file=temp/timeseries/timeseries_regression.log"])
    # print "log file: %s" % args.log_file
    configure_logger(args)

    dir_create("./temp/timeseries")  # for logging and plots

    random.seed(42)
    rnd.seed(42)

    reg_type = args.algo  # "nntf" # "nnsk" # "rfor" # "svr"
    n_anoms = args.n_anoms
    n_lags = args.n_lags
    n_epochs = args.n_epochs
    normalize_trend = args.normalize_trend
    batch_size = 20

    # datasets = univariate_timeseries_datasets.keys()
    dataset = args.dataset
    # dataset = "airline"
    logger.debug("dataset: %s, reg_type: %s" % (dataset, reg_type))
    data = get_univariate_timeseries_data(dataset)

    find_anomalies_with_regression(np.array(data, dtype=float), dataset, n_lags=n_lags,
                                   reg_type=reg_type, normalize_trend=normalize_trend,
                                   batch_size=batch_size,
                                   n_epochs=n_epochs, n_anoms=n_anoms)