def gen_a_periodic_shift(N, T, L, M, pA, k = 5, interval = [10,101], seed = None, noise_sig = 0.1, L2 = None, periods = None, periods2 = None):
""" Switches signal periods from one set of trends to another and back again
after L2 steps
Tests for a group of signals momentarily breaking correlations.
"""
if seed:
print 'Setting seed to %i' % (seed)
npr.seed(seed)
if periods is None:
periods = []
num_left = float(interval[1] - interval[0])
num_needed = float(k)
for i in xrange(interval[0], interval[1]+1):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
periods.append(i)
num_needed -= 1
else:
num_left -=1
if periods2 is None:
periods2 = []
num_left = float(interval[1] - interval[0])
num_needed = float(k)
for i in xrange(interval[0], interval[1]+1):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
periods2.append(i)
num_needed -= 1
else:
num_left -=1
# Seed already set at start of function
A, sins1 = sin_rand_combo(N, T, periods, noise_scale = noise_sig)
B, sins2 = sin_rand_combo(N, T, periods2, noise_scale = noise_sig)
# Anomaly will occur (and finish) Between time points 50 and T - 10
start_point = npr.randint(50, T - L - L2 - 10)
# Select stream(s) to be anomalous
if type(pA) == int:
num_anom = pA
elif pA < 1.0:
num_anom = np.floor(pA * N)
if num_anom > 1:
anoms = []
num_left = float(N)
num_needed = float(num_anom)
for i in xrange(N):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
anoms.append(i)
num_needed -= 1
num_left -=1
else:
num_left -=1
A[start_point:start_point+L2,anoms] = B[start_point:start_point+L2,anoms]
else:
anoms = npr.randint(N)
A[start_point:start_point+L2,anoms] = B[start_point:start_point+L2,anoms]
''' Ground Truth Table '''
if anoms.__class__ == int:
anoms = [anoms]
gt_table = np.zeros(num_anom*2, dtype = [('start','i4'),('loc','i4'),('len','i4'),('mag','i4'),('type','a10')])
count = 0
for i, a in enumerate(anoms):
gt_table[i+count] = (start_point, a, 10, M, 'Break Trends')
gt_table[i+count+1] = (start_point+L2, a, 10, M, 'Make Trends')
count += 1
output = dict(data = A, gt = gt_table, trends = sins1, trends2 = sins2, periods = periods, periods2 = periods2)
return output
def gen_a_peak_dip(N, T, L, M, pA, k = 5, interval = [10,121], seed = None, noise_sig = 0.1, L2 = None, periods = None, periods2 = None):
""" Adds short peak or dip to data
interval used is (low (Inclusive), High (exclusive)]
"""
if seed:
print 'Setting seed to %i' % (seed)
npr.seed(seed)
if periods is None:
periods = []
num_left = float(interval[1] - interval[0])
num_needed = float(k)
for i in xrange(interval[0], interval[1]+1):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
periods.append(i)
num_needed -= 1
else:
num_left -=1
#if seed is not None:
## Code to make random sins combination
#A, sins = sin_rand_combo(N, T, periods, seed = seed, noise_scale = noise_sig)
#else:
## Code to make random sins combination
# Seed already set at start of function
A, sins = sin_rand_combo(N, T, periods, noise_scale = noise_sig)
# Anomaly will occur (and finish) Between time points 50 and T - 10
start_point = npr.randint(50, T - L - 10)
# Code to make linear Anomalous trend
baseLine = 0
# Randomly choose peak or dip
if npr.rand() < 0.5:
anom_type = 'Peak'
s0lin = Tseries(0)
s0lin.makeSeries([1,3,4,1], [start_point, L/2, L/2, T - start_point - L],
[baseLine, baseLine, baseLine + M, baseLine],
gradient = float(M)/float(L/2), noise_type ='none')
else:
anom_type = 'Dip'
s0lin = Tseries(0)
s0lin.makeSeries([1,4,3,1], [start_point, L/2, L/2, T - start_point - L],
[baseLine, baseLine, baseLine - M, baseLine],
gradient = float(M)/float(L/2), noise_type ='none')
# Select stream(s) to be anomalous
if type(pA) == int:
num_anom = pA
elif pA < 1.0:
num_anom = np.floor(pA * N)
if num_anom > 1:
anoms = []
num_left = float(N)
num_needed = float(num_anom)
for i in xrange(N):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
anoms.append(i)
num_needed -= 1
num_left -= 1
else:
num_left -= 1
A[:,anoms] = A[:,anoms] + np.atleast_2d(s0lin).T
else:
anoms = npr.randint(N)
A[:,anoms] = A[:,anoms] + np.array(s0lin)
''' Grount truth Table '''
if anoms.__class__ == int:
anoms = [anoms]
gt_table = np.zeros(num_anom, dtype = [('start','i4'),('loc','i4'),('len','i4'),('mag','i4'),('type','a10')])
for i, a in enumerate(anoms):
gt_table[i] = (start_point, a, L, M, anom_type)
output = dict(data = A, gt = gt_table, trends = sins, periods = periods)
return output
def gen_a_grad_persist(N, T, L, M, pA, k = 5, interval = [10,101], seed = None, noise_sig = 0.1, L2 = None, periods = None, periods2 = None):
""" Adds longer persisted anomaly. gradient up/down to it
and vice versa after L steps """
if seed:
print 'Setting seed to %i' % (seed)
npr.seed(seed)
if periods is None:
periods = []
num_left = float(interval[1] - interval[0])
num_needed = float(k)
for i in xrange(interval[0], interval[1]+1):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
periods.append(i)
num_needed -= 1
else:
num_left -=1
# Seed already set at start of function
A, sins = sin_rand_combo(N, T, periods, noise_scale = noise_sig)
# Anomaly will occur (and finish) Between time points 50 and T - 10
start_point = npr.randint(50, T - L - L2 - 10)
# Code to make linear Anomalous trend
baseLine = 0
# Randomly choose peak or dip
if npr.rand() < 0.5:
anom_type = 'Grad Up--Down'
s0lin = Tseries(0)
s0lin.makeSeries([1,3,1,4,1], [start_point, L/2, L2, L/2, T - start_point - L - L2],
[baseLine, baseLine, baseLine + M, baseLine + M, baseLine],
gradient = float(M)/float(L/2), noise_type ='none')
else:
anom_type = 'Grad Down--Up'
s0lin = Tseries(0)
s0lin.makeSeries([1,4,1,3,1], [start_point, L/2, L2, L/2, T - start_point - L - L2],
[baseLine, baseLine, baseLine - M, baseLine - M, baseLine],
gradient = float(M)/float(L/2), noise_type ='none')
# Select stream(s) to be anomalous
if type(pA) == int:
num_anom = pA
elif pA < 1.0:
num_anom = np.floor(pA * N)
if num_anom > 1:
anoms = []
num_left = float(N)
num_needed = float(num_anom)
for i in xrange(N):
# probability of selection = (number needed)/(number left)
p = num_needed / num_left
if npr.rand() <= p:
anoms.append(i)
num_needed -= 1
num_left -=1
else:
num_left -=1
A[:,anoms] = A[:,anoms] + np.atleast_2d(s0lin).T
else:
anoms = npr.randint(N)
A[:,anoms] = A[:,anoms] + np.array(s0lin)
''' Ground Truth Table '''
if anoms.__class__ == int:
anoms = [anoms]
gt_table = np.zeros(num_anom*2, dtype = [('start','i4'),('loc','i4'),('len','i4'),('mag','i4'),('type','a10')])
count = 0
for i, a in enumerate(anoms):
a1_type = anom_type[5:].split('--')[0]
a2_type = anom_type[5:].split('--')[1]
gt_table[i+count] = (start_point, a, L, M, 'Grad ' + a1_type)
gt_table[i+count+1] = (start_point+L2+(L/2.0), a, L, M, 'Grad ' + a2_type)
count += 1
output = dict(data = A, gt = gt_table, trends = sins, periods = periods)
return output