-
Notifications
You must be signed in to change notification settings - Fork 2
/
Generate_data.py
255 lines (204 loc) · 9.55 KB
/
Generate_data.py
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import numpy as np
import matplotlib.pyplot as plt
from copy import deepcopy as copy
from tick.hawkes import (SimuHawkes, HawkesKernelTimeFunc, HawkesKernelExp, HawkesEM)
from tick.base import TimeFunction
from tick.plot import plot_hawkes_kernels
plt.rcParams['pdf.fonttype'] = 42
plt.rcParams['font.family'] = 'Calibri'
seed = 1111
np.random.seed(seed)
def gaussian(x, mu, sig):
return (np.exp(-np.power(x - mu, 2.) / (2 * np.power(sig, 2.))))/(2 * np.pi * np.power(sig, 2.)) ** 0.5
def kernel(dt, means, sigs, alpha):
k = gaussian(dt[:, None], means[None, :], sigs[None, :]).dot(alpha)
return k
def simulHawkes(lamb0, alpha, means, sigs, run_time=1000):
maxdt = max(means)+3*max(sigs)
nbClasses = len(alpha)
# Definition kernels
emptyKer = HawkesKernelTimeFunc(TimeFunction(([0, 1], [0, 0]), inter_mode=TimeFunction.InterConstRight))
kernels = [[copy(emptyKer) for _ in range(nbClasses)] for _ in range(nbClasses)]
for c in range(nbClasses):
for c2 in range(nbClasses):
if c!=c2: continue # Univariate Hawkes process
t_values = np.linspace(0, maxdt, 100)
y_values = kernel(t_values, means, sigs, alpha[c,c2])
#plt.plot(t_values, y_values)
#plt.show()
tf = TimeFunction((t_values, y_values), inter_mode=TimeFunction.InterConstRight, dt=maxdt/100)
#kernels.append(HawkesKernelTimeFunc(tf))
kernels[c][c2] = HawkesKernelTimeFunc(tf)
baseline = np.array([lamb0 for _ in range(nbClasses)])
hawkes = SimuHawkes(baseline=baseline, end_time=run_time, verbose=False, seed=int(np.random.random()*10000))
for c in range(nbClasses):
for c2 in range(nbClasses):
if c!=c2: continue # Univariate Hawkes process
hawkes.set_kernel(c, c2, kernels[c][c2])
hawkes.simulate()
events = []
for c, _ in enumerate(baseline):
events.append([c,0])
for c in range(len(hawkes.timestamps)):
for t in hawkes.timestamps[c]:
events.append([c, t])
events = np.array(events)
return events, hawkes
def simulTxt(events, voc_per_class, nbClasses, overlap_voc, words_per_obs):
# Generate text
# Perfectly separated text content
voc_clusters = [np.array(list(range(int(voc_per_class)))) + c*voc_per_class for c in range(nbClasses)]
# Overlap
for c in range(nbClasses):
voc_clusters[c] -= int(c*voc_per_class*overlap_voc)
# Associate a fraction of vocabulary to each observation
arrtxt = []
for e in events:
c_text = int(e[1])
arrtxt.append(np.random.choice(voc_clusters[c_text], size=words_per_obs))
return arrtxt
def compute_kernel_alltimes(events, means, sigs, alpha, res=1000):
ranget = np.linspace(0, np.max(events[:, -1]), res)
tabvals = [[] for _ in range(nbClasses)]
maxdt = max(means)+3*max(sigs)
for t in ranget:
ev = events[events[:, -1]>t-maxdt]
ev = ev[ev[:, -1]<t]
for c in range(nbClasses):
eventsprec = ev[ev[:, 0] == c] # 0 bc has to be temporal clusters
val = kernel(t - eventsprec[:, -1], means, sigs, alpha[c,c]).sum()
tabvals[c].append(val)
tabvals = np.array(tabvals)
return ranget, tabvals[0], tabvals[1]
def compute_overlap_temp(x, y1, y2):
area1 = np.trapz(y1, x=x)
area2 = np.trapz(y2, x=x)
areaInter = np.trapz(np.min([y1,y2], axis=0), x=x)
overlap = 2*areaInter/(area1+area2)
return overlap
def make_overlap_temp(events, alpha, overlap_temp, params_resimul):
ol_temp = -1000
eps = 0.05
dt = 10
res = 1000
while not (ol_temp>overlap_temp-eps and ol_temp<overlap_temp+eps):
maxt = np.max(events[:, -1])
i=0
while not (ol_temp>overlap_temp-eps and ol_temp<overlap_temp+eps) and nbClasses == 2 and i*dt<maxt:
events[events[:, 0]==0, 1] += dt
t, kernel1, kernel2 = compute_kernel_alltimes(events, means, sigs, alpha, res=res)
ol_temp = compute_overlap_temp(t, kernel1, kernel2)
i+=1
#print(i*dt, ol_temp, maxt)
if ol_temp<overlap_temp:
break
if not (ol_temp>overlap_temp-eps and ol_temp<overlap_temp+eps):
events, hawkes = simulHawkes(*params_resimul)
return events
def save(folder, name, events, arrtxt, lamb0, means, sigs, alpha):
events = np.insert(events, 3, np.array(list(range(len(events)))), axis=1)
events = np.array(list(sorted(events, key= lambda x: x[2])))
with open(folder+name+"_events.txt", "w+") as f:
for i, e in enumerate(events):
content = ",".join(map(str, list(arrtxt[int(e[3])])))
txt = str(e[0])+"\t"+str(e[1])+"\t"+str(e[2])+"\t"+content+"\n"
f.write(txt)
with open(folder+name+"_lamb0.txt", "w+") as f:
f.write(str(lamb0))
np.savetxt(folder+name+"_means.txt", means)
np.savetxt(folder+name+"_sigs.txt", sigs)
np.save(folder+name+"_alpha", alpha)
def plotProcess(events, means, sigs, alpha, whichclus=0):
colors = ["r", "b", "y", "g", "orange", "cyan","purple"]
maxdt = max(means)+3*max(sigs)
nbClasses = len(alpha)
rangedt = np.linspace(0, maxdt, 100)
for e in events:
c = int(e[whichclus])
t = e[-1]
#plt.plot(t, -1/10-c/10, "o", c=colors[c], markersize=4)
#plt.plot(t+rangedt, kernel(rangedt, means, sigs, alpha[c]), colors[c], alpha=0.1)
ranget = np.linspace(0, np.max(events[:, -1]), 10000)
tabvals = [[] for _ in range(nbClasses)]
for t in ranget:
ev = events[events[:, -1]>t-maxdt]
ev = ev[ev[:, -1]<t]
for c in range(nbClasses):
eventsprec = ev[ev[:, 0] == c] # 0 bc has to be temporal clusters
val = kernel(t - eventsprec[:, -1], means, sigs, alpha[c,c]).sum()
tabvals[c].append(val)
tabvals = np.array(tabvals)
for c in range(nbClasses):
plt.plot(ranget, lamb0+tabvals[c], "-", c=colors[c])
plt.show()
def generate(params):
nbClasses, run_time, voc_per_class, overlap_voc, overlap_temp, voc_per_class, perc_rand, words_per_obs, run, lamb0, means, sigs, folder = params
maxdt = max(means)+3*max(sigs)
alpha = np.zeros((nbClasses, nbClasses, len(means)))
for c in range(nbClasses):
a = np.random.random((len(means)))
alpha[c,c]=a/np.sum(a)
for c2 in range(nbClasses):
if c==c2: continue
a = np.random.random((len(means)))
alpha[c,c2] = 0
alpha = np.array(alpha)
# Get timestamps and temporal clusters
events = []
events, hawkes = simulHawkes(lamb0, alpha, means, sigs, run_time=run_time)
print(len(events), "events")
dofit = False
if dofit:
em = HawkesEM(15, kernel_size=30, n_threads=8, verbose=False, tol=1e-3)
em.fit(hawkes.timestamps)
fig = plot_hawkes_kernels(em, hawkes=hawkes, show=False)
plt.show()
# Get the wanted temporal overlap
if overlap_temp >=0 and nbClasses==2:
params_resimul=(lamb0, alpha, means, sigs, run_time)
events = make_overlap_temp(events, alpha, overlap_temp, params_resimul)
# Initialize textual clusters and shuffle nb_rand of them
events = np.insert(events, 0, events[:, 0], axis=1)
nb_rand = int(perc_rand*len(events))
events[np.random.randint(0, len(events), nb_rand), 1] = np.random.randint(0, nbClasses, nb_rand)
# Generate text associated with textual clusters
arrtxt = simulTxt(events, voc_per_class, nbClasses, overlap_voc, words_per_obs)
# Plot the process (univariate only e.g. diagonal of alpha)
#print(len(events))
#plotProcess(events, means, sigs, alpha, whichclus=1)
#pause()
name = f"Obs_nbclasses={nbClasses}_lg={run_time}_overlapvoc={overlap_voc}_overlaptemp={overlap_temp}_percrandomizedclus={perc_rand}_vocperclass={voc_per_class}_wordsperevent={words_per_obs}_run={run}"
save(folder, name, events, arrtxt, lamb0, means, sigs, alpha)
nbClasses = 2
run_time = 1500
XP = "Overlap"
overlap_voc = 0. # Proportion of voc in common between a clusters and its direct neighbours
overlap_temp = 0. # Overlap between the kernels of the simulating process
voc_per_class = 1000 # Number of words available for each cluster
perc_rand = 0. # Percentage of events to which assign random textual cluster
words_per_obs = 100
run = 0
lamb0 = 0.05
means = np.array([3, 7, 11])
sigs = np.array([0.5, 0.5, 0.5])
folder = "data/Synth/"
np.random.seed(1564)
#params = (nbClasses, run_time, voc_per_class, overlap_voc, overlap_temp, voc_per_class, perc_rand, words_per_obs, run, lamb0, means, sigs, folder)
params = (2, 60, voc_per_class, 0.3, 0.4, voc_per_class, perc_rand, words_per_obs, 0, lamb0, means, sigs, folder)
generate(params)
pause()
nbRuns = 10
if XP == "Decorr":
for perc_rand in np.array(list(range(11)))/10:
for run in range(nbRuns):
params = (nbClasses, run_time, voc_per_class, overlap_voc, overlap_temp, voc_per_class, perc_rand, words_per_obs, run, lamb0, means, sigs, folder)
print(f"{nbClasses} classes - OL_text={overlap_voc} - OL_temp={overlap_temp} - perc_rand={perc_rand} - run={run}")
generate(params)
elif XP == "Overlap":
np.random.seed(14776)
for overlap_voc in [0., 0.3, 0.5, 0.7, 0.9]:
for overlap_temp in [0., 0.3, 0.5, 0.7]:
for run in range(nbRuns):
params = (nbClasses, run_time, voc_per_class, overlap_voc, overlap_temp, voc_per_class, perc_rand, words_per_obs, run, lamb0, means, sigs, folder)
print(f"{nbClasses} classes - OL_text={overlap_voc} - OL_temp={overlap_temp} - perc_rand={perc_rand} - run={run}")
generate(params)