-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
338 lines (241 loc) · 11.8 KB
/
utils.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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
import os
import numpy as np
import matplotlib.pyplot as plt
import torch
from torch import nn
import pickle
from sklearn.metrics import accuracy_score, f1_score
# create the outdir
def create_outpath(dataset):
path = os.getcwd()
pid = os.getpid()
wsppath = os.path.join(path, 'workspace')
if not os.path.isdir(wsppath):
os.mkdir(wsppath)
outpath = os.path.join(wsppath, 'dataset:'+dataset + '-' + 'pid:'+str(pid))
assert not os.path.isdir(outpath), 'output directory already exist (process id coincidentally the same), please retry'
os.mkdir(outpath)
return outpath
def forward_pass(sde, eta0, time_seqs, type_seqs, mask, device, batch_size, dim_eta, num_divide, args):
padded_seq_length = len(time_seqs[0])
sde.batch_size = batch_size
eta_batch_l = torch.zeros(batch_size, padded_seq_length, dim_eta, device=device)
# eta_batch_l[i,j,:] is the left-limit of \mathbf{eta} at the j-th event in the i-th sequence
eta_batch_r = torch.zeros(batch_size, padded_seq_length, dim_eta, device=device)
eta_time_l = torch.zeros(batch_size, padded_seq_length, device=device)
# eta_time_l[i,j] == eta_batch_l[i,j,:][type_seqs[i,j]]
eta_time_r = torch.zeros(batch_size, padded_seq_length, device=device)
eta_batch_l[:, 0, :] = eta0.unsqueeze(0).repeat(batch_size,1)
event_type = type_seqs[:, 0].tolist()
eta_time_l[:, 0] = eta_batch_l[list(range(0, batch_size)), 0, event_type]
eta_batch_r[:, 0, :] = eta_batch_l[:, 0, :] + sde.h(eta_batch_l[:, 0, :].clone())[list(range(0, batch_size)), :, event_type]
eta_time_r[:, 0] = eta_batch_r[list(range(0, batch_size)), 0, event_type]
tsave = torch.Tensor().to(device)
eta_tsave = torch.Tensor().to(device)
eta_initial = eta_batch_r[:, 0, :]
for i in range(padded_seq_length-1):
adjacent_events = time_seqs[:, i:i+2]
ts, eta_ts_l = EulerSolver(sde, eta_initial, adjacent_events, num_divide, device)
tsave = torch.cat((tsave, ts), dim=1)
eta_tsave = torch.cat((eta_tsave, eta_ts_l), dim=2)
eta_batch_l[:, i+1, :] = eta_ts_l[:, :, -1]
eta_ts_r = eta_ts_l.clone()
event_type = type_seqs[:, i+1].tolist()
eta_ts_r[:, :, -1] = eta_ts_l[:, :, -1] + sde.h(eta_ts_l[:, :, -1])[list(range(0, batch_size)), :, event_type]
eta_batch_r[:, i+1, :] = eta_ts_r[:, :, -1]
eta_time_l[:, i+1] = eta_batch_l[list(range(0, batch_size)), i+1, event_type]
eta_time_r[:, i+1] = eta_batch_r[list(range(0, batch_size)), i+1, event_type]
eta_initial = eta_ts_r[:, :, -1]
masked_eta_time_l = eta_time_l * mask
sum_term = torch.sum(masked_eta_time_l)
mask_without_first_col = mask[:, 1:]
expanded_mask = mask_without_first_col.unsqueeze(2).repeat(1, 1, num_divide+1).view(mask.size(0), -1)
expanded_mask = expanded_mask.unsqueeze(1).repeat(1,dim_eta,1)
eta_tsave = eta_tsave * expanded_mask # mask the eta_tsave
expanded_diff_tsave = torch.diff(tsave).unsqueeze(1).repeat(1, dim_eta, 1)
integral_term = torch.sum((torch.exp(eta_tsave)[:, :, :-1] * expanded_mask[:, :, :-1] + torch.exp(eta_tsave)[:, :, 1:] * expanded_mask[:, :, 1:]) * (expanded_diff_tsave * expanded_mask[:, :, 1:])) / 2 # reason for mask: e^0=1
log_likelihood = sum_term - integral_term
loss = - log_likelihood
return eta_batch_l, eta_batch_r, loss
def EulerSolver(sde, eta_initial, adjacent_events, num_divide, device):
dt = torch.diff(adjacent_events, dim=1) / num_divide
ts = torch.cat([adjacent_events[:, 0].unsqueeze(dim=1) + dt * j for j in range(num_divide+1)], dim=1)
eta_ts = torch.Tensor().to(device)
eta_ts = torch.cat((eta_ts, eta_initial.unsqueeze(2)), dim=2)
for _ in range(num_divide):
eta_output = eta_initial + sde.f(eta_initial.clone())*dt + sde.g(eta_initial.clone())*torch.sqrt(dt)*torch.randn_like(eta_initial).to(device)
eta_ts = torch.cat((eta_ts, eta_output.unsqueeze(2)), dim=2)
eta_initial = eta_output
return ts, eta_ts
def next_predict(sde, eta0, time_seqs, type_seqs, device, dim_eta, num_divide, h=8, n_samples=1000, args=None):
"""
Predict the time and type of the next event given historical events.
"""
estimate_dt = []
next_dt = []
error_dt = []
estimate_type = []
next_type = []
loss_list = []
for idx, seq in enumerate(time_seqs):
seq_time = seq.unsqueeze(0)
seq_type = type_seqs[idx].unsqueeze(0)
mask = torch.ones(1, len(seq), device=device)
eta_seq_l, eta_seq_r, loss_idx = forward_pass(sde, eta0, seq_time, seq_type, mask, device, batch_size=1, dim_eta=dim_eta, num_divide=num_divide, args=args)
dt_seq = torch.diff(seq)
max_dt = torch.max(dt_seq)
estimate_seq_dt = []
next_seq_dt = dt_seq.tolist()
error_seq_dt = []
estimate_seq_type = []
next_seq_type = type_seqs[idx][1:].tolist()
for i in range(len(seq)-1):
last_t = seq[i]
n_dt = dt_seq[i]
timestep = h * max_dt / n_samples
tau = last_t + torch.linspace(0, h * max_dt, n_samples).to(device)
d_tau = tau - last_t
eta_last_t = eta_seq_r[:, i, :]
adjacent_events = torch.tensor([last_t, last_t+h * max_dt], device=device).unsqueeze(0)
_, eta_tau = EulerSolver(sde, eta_last_t, adjacent_events, n_samples-1, device)
eta_tau = eta_tau.squeeze(0)
intens_tau = torch.exp(eta_tau)
intens_tau_sum = intens_tau.sum(dim=0)
integral = torch.cumsum(timestep * intens_tau_sum, dim=0)
# density for the time-until-next-event law
density = intens_tau_sum * torch.exp(-integral)
d_tau_f_tau = d_tau * density
# trapezoidal method
e_dt = (timestep * 0.5 * (d_tau_f_tau[1:] + d_tau_f_tau[:-1])).sum()
err_dt = torch.abs(e_dt - n_dt)
e_type = torch.argmax(eta_seq_l[0][i+1])
estimate_seq_dt.append(e_dt.item())
error_seq_dt.append(err_dt.item())
estimate_seq_type.append(e_type.item())
loss_list.append(loss_idx)
estimate_dt.extend(estimate_seq_dt)
next_dt.extend(next_seq_dt)
error_dt.extend(error_seq_dt)
estimate_type.extend(estimate_seq_type)
next_type.extend(next_seq_type)
error_dt_tensor = torch.tensor(error_dt)
RMSE = np.linalg.norm(error_dt_tensor.detach().numpy(), ord=2) / (len(error_dt_tensor) ** 0.5)
acc = accuracy_score(next_type, estimate_type)
f1 = f1_score(next_type, estimate_type, average='weighted')
loss = sum(loss_list)
return loss, RMSE, acc, f1
def self_correcting_intensity(t, seq, mu, alpha):
intensity_val = np.exp(mu * t - alpha*len(seq[(seq>0) & (seq < t)]))
return intensity_val
def hawkes_intensity(t, seq, mu, alpha, beta):
intensity_val = mu + alpha * np.sum(np.exp(-beta * (t - seq[seq < t])))
return intensity_val
def poisson_intensity(t, seq, l):
intensity_val = l
return intensity_val
def recovery_intensity(sde, eta0, seq, device, num_divide):
t = torch.Tensor().to(device)
lmbda = torch.Tensor().to(device)
eta_initial = eta0.unsqueeze(0)
for i in range(len(seq)-1):
adjacent_events = seq[i:i+2].unsqueeze(0)
sde.batch_size = 1
ts, eta_ts_l = EulerSolver(sde, eta_initial, adjacent_events, num_divide, device)
t = torch.cat((t, ts))
lmbda = torch.cat((lmbda, torch.exp(eta_ts_l)))
eta_ts_r = eta_ts_l.clone()
eta_ts_r[0,0,-1] = eta_ts_l[0,0,-1] + sde.h(eta_ts_l[0,0,-1].unsqueeze(0))
eta_initial = eta_ts_r[:,:,-1]
return t, lmbda
def plot_poisson_recovery_intensity(t, lmbda, seq, xlabel, ylabel, title=''):
t = t.to('cpu')
t = t.detach().numpy()
lmbda = lmbda.to('cpu')
lmbda = lmbda.detach().numpy()
seq = seq.to('cpu')
seq = seq.detach().numpy()
true_poisson = [poisson_intensity(time, seq, l=1) for time in t]
plt.figure(figsize=(10, 4))
plt.plot(t, lmbda, linewidth=2, label=f'NJDTPP', color='#FF6666')
plt.plot(t, true_poisson, linewidth=2, label='ground truth', color='gray')
plt.scatter(seq, np.ones_like(seq) * min(true_poisson) * 0.05, marker='o', color='#1f77b4', label='events', s=5)
plt.title(title)
plt.xlabel(xlabel, fontsize=13)
plt.ylabel(ylabel, fontsize=13)
plt.legend()
plt.show()
def plot_hawkes_recovery_intensity(t, lmbda, seq, xlabel, ylabel, title=''):
t = t.to('cpu')
t = t.detach().numpy()
lmbda = lmbda.to('cpu')
lmbda = lmbda.detach().numpy()
seq = seq.to('cpu')
seq = seq.detach().numpy()
true_hawkes = [hawkes_intensity(time, seq, mu = 0.2, alpha = 0.8, beta = 1.0) for time in t]
plt.figure(figsize=(10, 4))
plt.plot(t, lmbda, linewidth=2, label=f'NJDTPP', color='#FF6666')
plt.plot(t, true_hawkes, linewidth=2, label='ground truth', color='gray')
plt.scatter(seq, np.ones_like(seq) * min(true_hawkes) * 0.05, marker='x', color='#1f77b4', label='events', s=5)
plt.title(title)
plt.xlabel(xlabel, fontsize=13)
plt.ylabel(ylabel, fontsize=13)
plt.legend()
plt.show()
def plot_self_correcting_recovery_intensity(t, lmbda, seq, xlabel, ylabel, title=''):
t = t.to('cpu')
t = t.detach().numpy()
lmbda = lmbda.to('cpu')
lmbda = lmbda.detach().numpy()
seq = seq.to('cpu')
seq = seq.detach().numpy()
true_self_correcting = [self_correcting_intensity(time, seq, mu=0.5, alpha=0.2) for time in t]
plt.figure(figsize=(10, 4))
plt.plot(t, lmbda, linewidth=2, label=f'NJDTPP', color='#FF6666')
plt.plot(t, true_self_correcting, linewidth=2, label='ground truth', color='gray')
plt.scatter(seq, np.ones_like(seq) * min(true_self_correcting) * 0.05, marker='o', color='#1f77b4', label='events', s=5)
plt.title(title)
plt.xlabel(xlabel, fontsize=13)
plt.ylabel(ylabel, fontsize=13)
plt.legend()
plt.show()
def load_pickle(file_dir):
"""Load from pickle file.
Args:
file_dir (BinaryIO): dir of the pickle file.
Returns:
any type: the loaded data.
"""
with open(file_dir, 'rb') as file:
try:
data = pickle.load(file, encoding='latin-1')
except Exception:
data = pickle.load(file)
return data
def build_input_from_pkl(device, source_dir: str, split: str):
"""
Args:
split (str, optional): denote the train, dev and test set.
"""
data = load_pickle(source_dir)
num_event_types = data["dim_process"]
source_data = data[split]
time_seqs = [[float(x["time_since_start"]) for x in seq] for seq in source_data if seq]
type_seqs = [[x["type_event"] for x in seq] for seq in source_data if seq]
type_seqs = [torch.tensor(type_seqs[i], device=device) for i in range(len(type_seqs))]
mins = [min(seq) for seq in time_seqs]
time_seqs = [[round(time - min_val, 6) for time in time_seq] for time_seq, min_val in zip(time_seqs, mins)]
time_seqs = [torch.tensor(time_seqs[i], device=device) for i in range(len(time_seqs))]
seqs_lengths = [torch.tensor(len(seq), device=device) for seq in time_seqs]
return time_seqs, type_seqs, num_event_types, seqs_lengths
def process_loaded_sequences(device, source_dir: str, split: str):
"""
Preprocess the dataset by padding the sequences.
"""
time_seqs, type_seqs, num_event_types, seqs_lengths = \
build_input_from_pkl(device, source_dir, split)
tmax = max([max(seq) for seq in time_seqs])
# Build a data tensor by padding
time_seqs = nn.utils.rnn.pad_sequence(time_seqs, batch_first=True, padding_value=tmax+1)
type_seqs = nn.utils.rnn.pad_sequence(type_seqs, batch_first=True, padding_value=0)
mask = (time_seqs != tmax+1).float()
return time_seqs, type_seqs, num_event_types, seqs_lengths, mask