def prepare_segment(func, dtype, datasetdscr):
segmentor = func.segmentor
func.activityFetcher.precompute(dtype)
procdata = Data(segmentor.__str__())
procdata.generator = segment(dtype, datasetdscr, segmentor)
procdata.set = []
procdata.label = []
procdata.set_window = []
procdata.acts = func.acts
procdata.s_events = dtype.s_events
procdata.a_events = dtype.a_events
i = 0
for x in procdata.generator:
if i % 10000 == 0:
logger.debug(segmentor.shortname(), i)
i += 1
procdata.set_window.append(x)
procdata.label.append(
func.activityFetcher.getActivity(dtype, x['window']))
del procdata.generator
procdata.label = np.array(procdata.label)
return procdata
def get_info(self):
func=self.functions
result=Data('Result')
result.shortrunname=func.shortrunname
result.functions={}
for f in func.__dict__:
obj = func.__dict__[f]
if isinstance(obj, MyTask):
result.functions[f]=(obj.shortname(),obj.params)
return result
def merge_split_overlap_IntervalTree(p_acts, r_acts):
tree = IntervalTree()
from result_analyse.visualisation import plotJoinTree
PACT = column_index(p_acts, 'Activity')
PSTIME = column_index(p_acts, 'StartTime')
PETIME = column_index(p_acts, 'EndTime')
for row in p_acts.values:
if (row[PACT] == 0):
continue
start = row[PSTIME]
end = row[PETIME]
startv = start.value
endv = end.value
if (startv == endv):
startv = startv - 1
#tree[start:end]={'P':{'Activitiy':act.Activity,'Type':'P','Data':act}]
d = Data('P-act')
d.P = {'Activity': row[PACT], 'StartTime': start, 'EndTime': end}
d.R = None
tree[startv:endv] = d
RACT = column_index(r_acts, 'Activity')
RSTIME = column_index(r_acts, 'StartTime')
RETIME = column_index(r_acts, 'EndTime')
for row in r_acts.values:
if (row[RACT] == 0):
continue
start = row[RSTIME]
end = row[RETIME]
startv = start.value
endv = end.value
if (startv == endv):
startv = startv - 1
#tree[start:end]=[{'Activitiy':act.Activity,'Type':'R','Data':act}]
d = Data('R-act')
d.P = None
d.R = {'Activity': row[RACT], 'StartTime': start, 'EndTime': end}
tree[startv:endv] = d
# cmTreePlot(tree)
tree.split_overlaps()
# cmTreePlot(tree)
def data_reducer(x, y):
res = Data('merge')
res.R = x.R
res.P = x.P
if not (y.P is None):
if (res.P is None) or y.P['EndTime'] < res.P['EndTime']:
res.P = y.P
if not (y.R is None):
if (res.R is None) or y.R['EndTime'] < res.R['EndTime']:
res.R = y.R
return res
tree.merge_equals(data_reducer=data_reducer)
return tree
def fewDaysSplit(self, data, count):
sensor_events = data.s_events
activity_events = data.a_events
sdate = sensor_events.time.apply(lambda x: x.date())
adate = activity_events.StartTime.apply(lambda x: x.date())
days = adate.unique()
import random
selecteddays = random.sample(list(days), count)
Train0 = Data('train_random_days' + str(selecteddays))
Train0.s_events = sensor_events.loc[sdate.isin(selecteddays)]
Train0.a_events = activity_events.loc[adate.isin(selecteddays)]
Train0.s_event_list = Train0.s_events.values
return Train0
def makeTrainTest(self, sensor_events, activity_events):
dataset_split = min(activity_events.StartTime) + \
((max(activity_events.EndTime)-min(activity_events.StartTime))*4/5)
dataset_split = pd.to_datetime(dataset_split.date()) # day
Train = Data('train')
Test = Data('test')
Train.s_events = sensor_events[sensor_events.time < dataset_split]
Train.a_events = activity_events[
activity_events.EndTime < dataset_split]
Train.s_event_list = Train.s_events.values
Test.s_events = sensor_events[sensor_events.time >= dataset_split]
Test.a_events = activity_events[
activity_events.EndTime >= dataset_split]
Test.s_event_list = Test.s_events.values
return Train, Test
def process(self, datasetdscr, dataset):
removekeys = []
for sid, info in datasetdscr.sensor_desc.iterrows():
last = {}
last['value'] = -1000
if (info.Nominal | info.OnChange):
continue
xs = dataset.s_events.loc[dataset.s_events.SID == sid]
min = xs.value.min()
max = xs.value.max()
#print(min, max, max-min)
invalid_changes = (max - min) * 0.1
for key, event in xs.iterrows():
invalid_changes = event['value'] * .1
if abs(last['value'] - event['value']) < invalid_changes:
#print (event)
removekeys.append(key)
continue
last = event
# print(removekeys)
d = Data(dataset.name)
d.s_events = dataset.s_events.drop(removekeys)
d.a_events = dataset.a_events
d.s_event_list = d.s_events.values
d.acts = dataset.acts
d.act_map = dataset.act_map
return d
def makeFoldTrainTest(self, sensor_events, activity_events, fold):
sdate = sensor_events.time.apply(lambda x: x.date())
adate = activity_events.StartTime.apply(lambda x: x.date())
days = adate.unique()
kf = KFold(n_splits=fold)
kf.get_n_splits(days)
for j, (train_index, test_index) in enumerate(kf.split(days)):
Train0 = Data('train_fold_' + str(j))
Train0.s_events = sensor_events.loc[sdate.isin(days[train_index])]
Train0.a_events = activity_events.loc[adate.isin(
days[train_index])]
Train0.s_event_list = Train0.s_events.values
Test0 = Data('test_fold_' + str(j))
Test0.s_events = sensor_events.loc[sdate.isin(days[test_index])]
Test0.a_events = activity_events.loc[adate.isin(days[test_index])]
Test0.s_event_list = Test0.s_events.values
yield Train0, Test0
def load():
gt = [(65, 141), (157, 187), (260, 304), (324, 326), (380, 393),
(455, 470), (475, 485), (505, 555), (666, 807), (814, 888),
(903, 929)]
a = [
(66, 73), (78, 126), (135, 147), (175, 186), (225, 236), (274, 318),
(349, 354), (366, 372), (423, 436), (453, 460), (467, 473), (487, 493),
(501, 506), (515, 525), (531, 542), (545, 563), (576, 580), (607, 611),
(641, 646), (665, 673), (678, 898), (907, 933)
]
b = [(63, 136), (166, 188), (257, 310), (451, 473), (519, 546), (663, 916)]
dataset = Data('test')
dataset.activities = ['null', 'Act']
dataset.activities_map = {0: 'null', 1: 'Act'}
dataset.activities_map_inverse = {'null': 0, 'Act': 1}
dataset.activity_events = pd.DataFrame(
columns=["StartTime", "EndTime", "Activity", 'Duration'])
init = pd.to_datetime('1/1/2020')
for k in gt:
actevent = {
"StartTime": init + pd.to_timedelta(str(k[0]) + 's'),
"EndTime": init + pd.to_timedelta(str(k[1]) + 's'),
"Activity": 1,
'Duration': pd.to_timedelta(str(k[1] - k[0]) + 's')
}
dataset.activity_events = dataset.activity_events.append(
actevent, ignore_index=True)
aevents = pd.DataFrame(
columns=["StartTime", "EndTime", "Activity", 'Duration'])
for k in a:
actevent = {
"StartTime": init + pd.to_timedelta(str(k[0]) + 's'),
"EndTime": init + pd.to_timedelta(str(k[1]) + 's'),
"Activity": 1,
'Duration': pd.to_timedelta(str(k[1] - k[0]) + 's')
}
aevents = aevents.append(actevent, ignore_index=True)
bevents = pd.DataFrame(
columns=["StartTime", "EndTime", "Activity", 'Duration'])
for k in b:
actevent = {
"StartTime": init + pd.to_timedelta(str(k[0]) + 's'),
"EndTime": init + pd.to_timedelta(str(k[1]) + 's'),
"Activity": 1,
'Duration': pd.to_timedelta(str(k[1] - k[0]) + 's')
}
bevents = bevents.append(actevent, ignore_index=True)
return dataset, aevents, bevents
def data_reducer(x, y):
res = Data('merge')
res.R = x.R
res.P = x.P
if not (y.P is None):
if (res.P is None) or y.P['EndTime'] < res.P['EndTime']:
res.P = y.P
if not (y.R is None):
if (res.R is None) or y.R['EndTime'] < res.R['EndTime']:
res.R = y.R
return res
def merge_split_overlap_IntervalTree(p_acts, r_acts):
tree = IntervalTree()
for act in p_acts:
if (act['Activity'] == 0):
continue
start = act['StartTime'].value
end = act['EndTime'].value
if (start == end):
start = start - 1
#tree[start:end]={'P':{'Activitiy':act.Activity,'Type':'P','Data':act}]
d = Data('P-act')
d.P = act
d.R = None
tree[start:end] = d #{'P':act,'PActivitiy':act.Activity}
for act in r_acts:
start = act['StartTime'].value
end = act['EndTime'].value
if (start == end):
start = start - 1
#tree[start:end]=[{'Activitiy':act.Activity,'Type':'R','Data':act}]
d = Data('P-act')
d.P = None
d.R = act
tree[start:end] = d #{'R':act,'RActivitiy':act.Activity}
tree.split_overlaps()
def data_reducer(x, y):
res = x
if not (y.P is None):
if (res.P is None) or y.P['EndTime'] < res.P['EndTime']:
res.P = y.P
if not (y.R is None):
if (res.R is None) or y.R['EndTime'] < res.R['EndTime']:
res.R = y.R
return res
tree.merge_equals(data_reducer=data_reducer)
return tree
def justifySet(self, acts, Train, add_other_action=False):
inp = [Train]
out = []
if (acts[0] != 0):
acts.insert(0, 0)
act_map = {a: i for i, a in enumerate(acts)}
for dtype in inp:
ndtype = Data(dtype.name)
ndtype.s_events = dtype.s_events
ndtype.a_events = dtype.a_events.copy()
ndtype.a_events.Activity = dtype.a_events.Activity.apply(
lambda x: act_map[x]
if x in act_map else (len(acts) if add_other_action else 0))
out.append(ndtype)
ndtype.act_map = act_map
ndtype.acts = acts
return out[0] #Train
def fusion(self, results, real_events, isTrain):
intree = IntervalTree()
logger.info("\n=======================fusion activties ========")
# intree = IntervalTree()
# Segmentaion ###########################
for indx, tacts in enumerate(self.gacts):
result = results[indx]
for i in range(0, len(result.Sdata.set_window)):
idx = result.Sdata.set_window[i]
start = result.Sdata.s_event_list[idx[0], 1]
end = result.Sdata.s_event_list[idx[-1], 1]
rcls = result.Sdata.label[i]
pcls = result.predicted_classes[i]
fullprob = result.predicted[i]
if (end == start):
continue
d = Data(str(i))
d.real = rcls
d.pred = pcls
d.pred_prob = fullprob
if (isTrain):
self.train_quality[indx] = result.quality
d.gindx = indx
# {'real':rcls,'pred':pcls,'pred_prob':fullprob,'train_q':result.quality}
intree[start:end] = d
intree.split_overlaps()
segments = defaultdict(dict)
for item in intree.items():
segments[item.begin.value << 64
| item.end.value]['begin'] = item.begin
segments[item.begin.value << 64 | item.end.value]['end'] = item.end
segments[item.begin.value << 64
| item.end.value][item.data.gindx] = item.data
# Feature Extraction ###########################
f = np.zeros((len(segments), len(self.gacts) * len(self.acts)))
label = np.zeros(len(segments))
times = []
iseg = 0
for timeseg in segments:
seg = segments[timeseg]
b = seg['begin']
e = seg['end']
times.append({'begin': b, 'end': e})
for indx in range(len(self.gacts)):
if (indx in seg):
label[iseg] = seg[indx].real
start = indx * len(self.acts)
end = (indx + 1) * len(self.acts)
if (self.train_quality[indx]['f1'] < 0.1):
continue
f[iseg, start:end] = seg[indx].pred_prob
iseg += 1
#TRAIN #######################
if (isTrain):
inputsize = (len(f[0]), )
outputsize = len(self.acts)
self.fusion_model = tf.keras.models.Sequential([
tf.keras.layers.Dense(128, input_shape=inputsize),
tf.keras.layers.Dense(512, activation=tf.nn.relu),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(outputsize, activation=tf.nn.softmax)
],
name='fusion')
if (np.max(label) == 0):
# self.trained=False
cw = np.ones(len(self.acts))
else:
cw = compute_class_weight("balanced", self.acts, label)
self.fusion_model.summary()
self.fusion_model.compile(
optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=[
tf.keras.metrics.SparseCategoricalAccuracy(name='acc')
])
self.fusion_model.fit(f, label, epochs=10, class_weight=cw)
#EVALUATE #######################
result = Data('result')
result.results = results
result.predicted = self.fusion_model.predict(f)
result.predicted_classes = self.fusion_model.predict_classes(f)
# predicted = np.argmax(model.predict(f), axis=1)
pred_events = []
ptree = {}
epsilon = pd.to_timedelta('1s')
for i in range(len(f)):
start = times[i]['begin']
end = times[i]['end']
pclass = result.predicted_classes[i]
pred_events.append({
'Activity': pclass,
'StartTime': start,
'EndTime': end
})
pred_events = pd.DataFrame(pred_events)
pred_events = pred_events.sort_values(['StartTime'])
pred_events = pred_events.reset_index()
pred_events = pred_events.drop(['index'], axis=1)
result.shortrunname = "fusion model" + str(
{r: results[r].shortrunname
for r in results})
result.times = times
result.pred_events = pred_events
result.real_events = real_events
result.event_cm = event_confusion_matrix(result.real_events,
result.pred_events, self.acts)
result.quality = CMbasedMetric(result.event_cm, 'macro')
result.functions = {r: results[r].functions for r in results}
logger.debug('Evalution quality is %s' % result.quality)
return result
import feature_extraction.Raw
# from general.libimport import *
from general.utils import Data
import metric.classical
import ml_strategy.Simple
import ml_strategy.FastFinder
import ml_strategy.SeperateGroup
import ml_strategy.WeightedGroup
import ml_strategy.WeightedGroup2
import preprocessing.SimplePreprocessing
import segmentation.Probabilistic
import segmentation.FixedEventWindow
import segmentation.FixedSlidingWindow
import segmentation.FixedTimeWindow
methods = Data('methods')
methods.segmentation = [
{'method': lambda: segmentation.FixedEventWindow.FixedEventWindow(), 'params': [
{'var': 'size', 'min': 10, 'max': 30, 'type': 'int', 'init': 10, 'range':list(range(10,26,5))},
{'var': 'shift', 'min': 2, 'max': 20, 'type': 'int', 'init': 10, 'range':list(range(10,16,5))}
], 'findopt': False},
{'method': lambda: segmentation.FixedSlidingWindow.FixedSlidingWindow(), 'params': [
{'var': 'size' , 'min': 60, 'max': 15*60, 'type': 'float', 'init': 120/4, 'range':list(range(15,76,15))},
{'var': 'shift', 'min': 10, 'max': 7*60 , 'type': 'float', 'init': 60/2, 'range':list(range(15,45,15))}
], 'findopt': False},
{'method': lambda: segmentation.Probabilistic.Probabilistic(), 'params': [], 'findopt':False},
# {'method': lambda:segmentation.FixedTimeWindow.FixedTimeWindow(), 'params':[
# {'var':'size','min':pd.Timedelta(1, unit='s').total_seconds(), 'max': pd.Timedelta(30, unit='m').total_seconds(), 'type':'float','init':pd.Timedelta(15, unit='s').total_seconds()},
# {'var':'shift','min':pd.Timedelta(1, unit='s').total_seconds(), 'max': pd.Timedelta(30, unit='m').total_seconds(), 'type':'float','init':pd.Timedelta(1, unit='s').total_seconds()}
def fusion(self, results, real_events, isTrain):
intree = IntervalTree()
logger.info("\n=======================fusion activties ========")
# intree = IntervalTree()
# Segmentaion ###########################
for indx, tacts in enumerate(self.gacts):
result = results[indx]
for i in range(0, len(result.Sdata.set_window)):
idx = result.Sdata.set_window[i]
start = result.Sdata.s_event_list[idx[0], 1]
end = result.Sdata.s_event_list[idx[-1], 1]
rcls = result.Sdata.label[i]
pcls = result.predicted_classes[i]
fullprob = result.predicted[i]
if (end == start):
continue
d = Data(str(i))
d.real = rcls
d.pred = pcls
d.pred_prob = fullprob
if (isTrain):
self.train_quality[indx] = result.quality
d.gindx = indx
# {'real':rcls,'pred':pcls,'pred_prob':fullprob,'train_q':result.quality}
intree[start:end] = d
intree.split_overlaps()
segments = defaultdict(dict)
for item in intree.items():
segments[item.begin.value << 64
| item.end.value]['begin'] = item.begin
segments[item.begin.value << 64 | item.end.value]['end'] = item.end
segments[item.begin.value << 64
| item.end.value][item.data.gindx] = item.data
probs = np.zeros((len(segments), len(self.acts)))
# Feature Extraction ###########################
label = np.zeros(len(segments))
times = []
iseg = 0
for timeseg in segments:
seg = segments[timeseg]
b = seg['begin']
e = seg['end']
times.append({'begin': b, 'end': e})
for indx in range(len(self.gacts)):
if (indx in seg):
label[iseg] = seg[indx].real
if (self.mode == 1):
probs[iseg, :] += np.array(
seg[indx].pred_prob
) * self.train_quality[indx]['f1'] / len(self.gacts)
elif self.mode == 2:
p = np.zeros(len(self.acts))
p[np.argmax(seg[indx].pred_prob)] = 1
probs[iseg, :] += p
else:
p = np.zeros(len(self.acts))
p[np.argmax(seg[indx].pred_prob
)] = self.train_quality[indx]['f1']
probs[iseg, :] += p
iseg += 1
plabel = np.argmax(probs, 1)
#EVALUATE #######################
result = Data('result')
result.results = results
result.predicted = probs
result.predicted_classes = plabel
# predicted = np.argmax(model.predict(f), axis=1)
pred_events = []
ptree = {}
epsilon = pd.to_timedelta('1s')
for i in range(len(segments)):
start = times[i]['begin']
end = times[i]['end']
pclass = result.predicted_classes[i]
pred_events.append({
'Activity': pclass,
'StartTime': start,
'EndTime': end
})
pred_events = pd.DataFrame(pred_events)
pred_events = pred_events.sort_values(['StartTime'])
pred_events = pred_events.reset_index()
pred_events = pred_events.drop(['index'], axis=1)
result.shortrunname = "fusion model" + str(
{r: results[r].shortrunname
for r in results})
result.times = times
result.pred_events = pred_events
result.real_events = real_events
result.event_cm = event_confusion_matrix(result.real_events,
result.pred_events, self.acts)
result.quality = CMbasedMetric(result.event_cm, 'macro')
result.functions = {r: results[r].functions for r in results}
logger.debug('Evalution quality is %s' % result.quality)
return result
def method_param_selector(callback, uniquekey):
import itertools
from constants import methods
s = [
methods.preprocessing, methods.segmentation, methods.activity_fetcher,
methods.feature_extraction, methods.classifier
]
permut = list(itertools.product(*s))
allpool = []
for item in permut:
func = Data('Functions')
func.uniquekey = uniquekey
func.preprocessor = createFunction(item[0])
func.segmentor = createFunction(item[1])
func.activityFetcher = createFunction(item[2])
func.featureExtractor = createFunction(item[3])
func.classifier = createFunction(item[4])
func.combiner = createFunction(methods.combiner[0])
func.classifier_metric = createFunction(methods.classifier_metric[0])
func.event_metric = createFunction(methods.classifier_metric[0])
func.shortrunname = ''
for k in func.__dict__:
obj = func.__dict__[k]
if isinstance(obj, MyTask):
obj.func = func
func.shortrunname += obj.shortname() + '_'
optl = OptLearn(func, callback)
allpool.append(optl)
# break
success, fail = run(allpool, True)
bestJobscore = success[0].result['optq']['q']
bestJob = success[0]
for job in success:
if (bestJobscore > job.result['optq']['q']):
bestJobscore = job.result['optq']['q']
bestJob = job
return bestJob
def pipeline(self,func,data,train):
import os
os.system("taskset -p 0xff %d" % os.getpid())
func.acts=self.acts
logger.debug('Starting .... %s' % (func.shortrunname))
Tdata=func.preprocessor.process(self.datasetdscr, data)
logger.debug('Preprocessing Finished %s' % (func.preprocessor.shortname()))
Sdata=prepare_segment2(func,Tdata,self.datasetdscr)
logger.debug('Segmentation Finished %d segment created %s' % (len(Sdata.set_window), func.segmentor.shortname()))
Sdata.set=featureExtraction(func.featureExtractor,self.datasetdscr,Sdata,True)
logger.debug('FeatureExtraction Finished shape %s , %s' % (str(Sdata.set.shape), func.featureExtractor.shortname()))
if(train):
func.classifier.createmodel(Sdata.set[0].shape,len(self.acts))
func.classifier.setWeight(self.weight)
logger.debug('Classifier model created %s' % (func.classifier.shortname()))
func.classifier.train(Sdata.set, Sdata.label)
logger.debug('Classifier model trained %s' % (func.classifier.shortname()))
logger.info("Evaluating....")
result=Data('Result')
result.shortrunname=func.shortrunname
result.Sdata=Sdata
result.functions={}
for f in func.__dict__:
obj = func.__dict__[f]
if isinstance(obj, MyTask):
result.functions[f]=(obj.shortname(),obj.params)
result.predicted =func.classifier.predict(Sdata.set)
result.predicted_classes=func.classifier.predict_classes(Sdata.set)
pred_events =func.combiner.combine(Sdata.s_event_list,Sdata.set_window,result.predicted)
logger.debug('events merged %s' % (func.combiner.shortname()))
result.pred_events =pred_events
result.real_events =data.a_events
result.event_cm =event_confusion_matrix(data.a_events,pred_events,self.acts)
result.quality =CMbasedMetric(result.event_cm,'macro',self.weight)
#eventeval=EventBasedMetric(Sdata.a_events,pred_events,self.acts)
logger.debug('Evalution quality is %s'%result.quality)
return result
model.evaluate(f,label)
combiner=EmptyCombiner()
predicted=model.predict_classes(f)
# predicted = np.argmax(model.predict(f), axis=1)
pred_events = []
ptree = {}
epsilon=pd.to_timedelta('1s')
for i in range(len(f)):
start = times[i][0]
end = times[i][1]
pclass = predicted[i]
pred_events.append({'Activity': pclass, 'StartTime': start, 'EndTime': end})
pred_events = pd.DataFrame(pred_events)
pred_events = pred_events.sort_values(['StartTime'])
pred_events = pred_events.reset_index()
pred_events = pred_events.drop(['index'], axis=1)
result=Data('result')
result.pred_events =pred_events
result.real_events =savedata.train_results[11].real_events
result.event_cm =event_confusion_matrix(result.real_events,result.pred_events,savedata.acts)
result.quality =CMbasedMetric(result.event_cm,'macro')
print(result.quality)
def create(real, pred, filename):
evalres = [{}]
evalres[0]['test'] = Data('test res')
evalres[0]['test'].real_events = vs.convert2event(real)
evalres[0]['test'].pred_events = vs.convert2event(pred)
evalres[0]['test'].quality = {}
dataset = Data('MyDataset')
dataset.activities = ['None', 'Act']
dataset.activity_events = evalres[0]['test'].real_events
dataset.activities_map_inverse = {
k: v
for v, k in enumerate(dataset.activities)
}
dataset.activities_map = {v: k for v, k in enumerate(dataset.activities)}
dataset.sensor_events = pd.DataFrame()
runinfo = filename
utils.saveState([runinfo, dataset, evalres], filename)
