|
1
|
|
|
import os |
|
2
|
|
|
import pickle |
|
3
|
|
|
import logging |
|
4
|
|
|
import argparse |
|
5
|
|
|
from sklearn.ensemble import RandomForestClassifier |
|
6
|
|
|
from datetime import datetime |
|
7
|
|
|
from pyActLearn.CASAS.data import CASASData |
|
8
|
|
|
from pyActLearn.CASAS.fuel import CASASFuel |
|
9
|
|
|
from pyActLearn.performance.record import LearningResult |
|
10
|
|
|
from pyActLearn.performance import get_confusion_matrix |
|
11
|
|
|
|
|
12
|
|
|
logger = logging.getLogger(__file__) |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
def training_and_test(token, train_data, test_data, num_classes, result): |
|
16
|
|
|
"""Train and test |
|
17
|
|
|
|
|
18
|
|
|
Args: |
|
19
|
|
|
token (:obj:`str`): token representing this run |
|
20
|
|
|
train_data (:obj:`tuple` of :obj:`numpy.array`): Tuple of training feature and label |
|
21
|
|
|
test_data (:obj:`tuple` of :obj:`numpy.array`): Tuple of testing feature and label |
|
22
|
|
|
num_classes (:obj:`int`): Number of classes |
|
23
|
|
|
result (:obj:`pyActLearn.performance.record.LearningResult`): LearningResult object to hold learning result |
|
24
|
|
|
""" |
|
25
|
|
|
model = RandomForestClassifier(n_estimators=20, criterion="entropy") |
|
26
|
|
|
model.fit(train_data[0], train_data[1].flatten()) |
|
27
|
|
|
# Test |
|
28
|
|
|
predicted_y = model.predict(test_data[0]) |
|
29
|
|
|
predicted_proba = model.predict_proba(test_data[0]) |
|
30
|
|
|
# Evaluate the Test and Store Result |
|
31
|
|
|
confusion_matrix = get_confusion_matrix(num_classes=num_classes, |
|
32
|
|
|
label=test_data[1].flatten(), predicted=predicted_y) |
|
33
|
|
|
result.add_record(model.get_params(), key=token, confusion_matrix=confusion_matrix) |
|
34
|
|
|
# In case any label is missing, populate it |
|
35
|
|
|
if predicted_proba.shape[1] != num_classes: |
|
36
|
|
|
temp_array = np.zeros((predicted_proba.shape[0], num_classes), np.float32) |
|
37
|
|
|
for i in range(len(model.classes_)): |
|
38
|
|
|
temp_array[:, model.classes_[i]] = predicted_proba[:, i] |
|
39
|
|
|
predicted_proba = temp_array |
|
40
|
|
|
return predicted_y, predicted_proba |
|
41
|
|
|
|
|
42
|
|
|
|
|
43
|
|
|
def load_and_test(token, test_data, num_classes, result): |
|
44
|
|
|
"""Load and test |
|
45
|
|
|
|
|
46
|
|
|
Args: |
|
47
|
|
|
token (:obj:`str`): token representing this run |
|
48
|
|
|
test_data (:obj:`tuple` of :obj:`numpy.array`): Tuple of testing feature and label |
|
49
|
|
|
num_classes (:obj:`int`): Number of classes |
|
50
|
|
|
result (:obj:`pyActLearn.performance.record.LearningResult`): LearningResult object to hold learning result |
|
51
|
|
|
""" |
|
52
|
|
|
model = RandomForestClassifier(n_estimators=20, criterion="entropy") |
|
53
|
|
|
model.set_params(result.get_record_by_key(token)['model']) |
|
54
|
|
|
# Test |
|
55
|
|
|
predicted_y = model.predict(test_data[0]) |
|
56
|
|
|
predicted_proba = model.predict_proba(test_data[0]) |
|
57
|
|
|
return predicted_y, predicted_proba |
|
58
|
|
|
|
|
59
|
|
View Code Duplication |
if __name__ == '__main__': |
|
|
|
|
|
|
60
|
|
|
args_ok = False |
|
61
|
|
|
parser = argparse.ArgumentParser(description='Run Decision Tree on single resident CASAS datasets.') |
|
62
|
|
|
parser.add_argument('-d', '--dataset', help='Directory to original datasets') |
|
63
|
|
|
parser.add_argument('-o', '--output', help='Output folder') |
|
64
|
|
|
parser.add_argument('--h5py', help='HDF5 dataset folder') |
|
65
|
|
|
args = parser.parse_args() |
|
66
|
|
|
# Default parameters |
|
67
|
|
|
log_filename = os.path.basename(__file__).split('.')[0] + \ |
|
68
|
|
|
'-%s.log' % datetime.now().strftime('%y%m%d_%H:%M:%S') |
|
69
|
|
|
# Setup output directory |
|
70
|
|
|
output_dir = args.output |
|
71
|
|
|
if output_dir is not None: |
|
72
|
|
|
output_dir = os.path.abspath(os.path.expanduser(output_dir)) |
|
73
|
|
|
if os.path.exists(output_dir): |
|
74
|
|
|
# Found output_dir, check if it is a directory |
|
75
|
|
|
if not os.path.isdir(output_dir): |
|
76
|
|
|
exit('Output directory %s is found, but not a directory. Abort.' % output_dir) |
|
77
|
|
|
else: |
|
78
|
|
|
# Create directory |
|
79
|
|
|
os.makedirs(output_dir) |
|
80
|
|
|
else: |
|
81
|
|
|
output_dir = '.' |
|
82
|
|
|
# If dataset is specified, update h5py |
|
83
|
|
|
casas_data_dir = args.dataset |
|
84
|
|
|
if casas_data_dir is not None: |
|
85
|
|
|
casas_data_dir = os.path.abspath(os.path.expanduser(casas_data_dir)) |
|
86
|
|
|
if not os.path.isdir(casas_data_dir): |
|
87
|
|
|
exit('CASAS dataset at %s does not exist. Abort.' % casas_data_dir) |
|
88
|
|
|
# Find h5py dataset first |
|
89
|
|
|
h5py_dir = args.h5py |
|
90
|
|
|
if h5py_dir is not None: |
|
91
|
|
|
h5py_dir = os.path.abspath(os.path.expanduser(h5py_dir)) |
|
92
|
|
|
else: |
|
93
|
|
|
# Default location |
|
94
|
|
|
h5py_dir = os.path.join(output_dir, 'h5py') |
|
95
|
|
|
if os.path.exists(h5py_dir): |
|
96
|
|
|
if not os.path.isdir(h5py_dir): |
|
97
|
|
|
exit('h5py dataset location %s is not a directory. Abort.' % h5py_dir) |
|
98
|
|
|
log_filename = os.path.join(output_dir, log_filename) |
|
99
|
|
|
# Setup Logging as early as possible |
|
100
|
|
|
logging.basicConfig(level=logging.DEBUG, |
|
101
|
|
|
format='[%(asctime)s] %(name)s:%(levelname)s:%(message)s', |
|
102
|
|
|
handlers=[logging.FileHandler(log_filename), |
|
103
|
|
|
logging.StreamHandler()]) |
|
104
|
|
|
if not CASASFuel.files_exist(h5py_dir): |
|
105
|
|
|
# Finish check and creating all directory needed - now load datasets |
|
106
|
|
|
if casas_data_dir is not None: |
|
107
|
|
|
casas_data = CASASData(path=casas_data_dir) |
|
108
|
|
|
casas_data.summary() |
|
109
|
|
|
# SVM needs to use statistical feature with per-sensor and normalization |
|
110
|
|
|
casas_data.populate_feature(method='stat', normalized=False, per_sensor=False) |
|
111
|
|
|
casas_data.export_hdf5(h5py_dir) |
|
112
|
|
|
casas_fuel = CASASFuel(dir_name=h5py_dir) |
|
113
|
|
|
# Prepare learning result |
|
114
|
|
|
result_pkl_file = os.path.join(output_dir, 'result.pkl') |
|
115
|
|
|
result = None |
|
116
|
|
|
if os.path.isfile(result_pkl_file): |
|
117
|
|
|
f = open(result_pkl_file, 'rb') |
|
118
|
|
|
result = pickle.load(f) |
|
119
|
|
|
f.close() |
|
120
|
|
|
if result.data != h5py_dir: |
|
121
|
|
|
logger.error('Result pickle file found for different dataset %s' % result.data) |
|
122
|
|
|
exit('Cannot save learning result at %s' % result_pkl_file) |
|
123
|
|
|
else: |
|
124
|
|
|
result = LearningResult(name='DecisionTree', data=h5py_dir, mode='by_week') |
|
125
|
|
|
num_classes = casas_fuel.get_output_dims() |
|
126
|
|
|
# Open Fuel and get all splits |
|
127
|
|
|
split_list = casas_fuel.get_set_list() |
|
128
|
|
|
train_names = ('week 24', 'week 23', 'week 22', 'week 21') |
|
129
|
|
|
test_names = ('week 25', 'week 26', 'week 27', 'week 28') |
|
130
|
|
|
test_name = 'single_test' |
|
131
|
|
|
train_set = casas_fuel.get_dataset(train_names, load_in_memory=True) |
|
132
|
|
|
(train_set_data) = train_set.data_sources |
|
133
|
|
|
test_set = casas_fuel.get_dataset(test_names, load_in_memory=True) |
|
134
|
|
|
(test_set_data) = test_set.data_sources |
|
135
|
|
|
# Prepare Back Annotation |
|
136
|
|
|
fp_back_annotated = open(os.path.join(output_dir, 'back_annotated.txt'), 'w') |
|
137
|
|
|
fp_back_probability = open(os.path.join(output_dir, 'back_annotated_proba.txt'), 'w') |
|
138
|
|
|
# run svm |
|
139
|
|
|
logger.info('Training on %s, Testing on %s' % (str(train_names), str(test_names))) |
|
140
|
|
|
if result.get_record_by_key(test_name) is None: |
|
141
|
|
|
prediction, prediction_proba = training_and_test(test_name, train_set_data, test_set_data, num_classes, result) |
|
142
|
|
|
else: |
|
143
|
|
|
prediction, prediction_proba = load_and_test(test_name, test_set_data, num_classes, result) |
|
144
|
|
|
casas_fuel.back_annotate(fp_back_annotated, prediction=prediction, split_name=test_names) |
|
145
|
|
|
casas_fuel.back_annotate_with_proba(fp_back_probability, prediction_proba=prediction_proba, split_name=test_names) |
|
146
|
|
|
train_name = test_name |
|
147
|
|
|
train_set_data = test_set_data |
|
148
|
|
|
f = open(result_pkl_file, 'wb') |
|
149
|
|
|
pickle.dump(obj=result, file=f, protocol=pickle.HIGHEST_PROTOCOL) |
|
150
|
|
|
f.close() |
|
151
|
|
|
result.export_to_xlsx(os.path.join(output_dir, 'result.xlsx')) |
|
152
|
|
|
|
|
153
|
|
|
|
TinghuiWang /
pyActLearn
Loading history...