-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
210 lines (168 loc) · 9.08 KB
/
model.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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import utils.constants as constants
from modules.HypUserEmbedding import HypUserEmbedding
from modules.RotaryLorentzAttention import LorentzSelfAttention
from utils.others import get_previous_user_mask
from manifolds.hyperboloid import Hyperboloid
from manifolds.lorentz_functions import givens_rotations, distance_prediction_scores
from modules.TimeModule import TimeModule
from utils.run_utils import make_type_mask_for_pad_sequence, compute_loglikelihood
class HHDiff(nn.Module):
def __init__(self, args):
super(HHDiff, self).__init__()
# parameters
self.device = args['device']
self.n_user = args['n_user']
config = args['model']
self.c = config['c']
self.dim = config['n_dim']
self.init_size = config['init_emb_size']
self.rotary_position_steps = config['rot_pe_steps']
self.dropout = config['dropout']
# Computing log-likelihood for the Hawkes prcoess
self.loss_integral_num_sample_per_step = config['n_sampled_times']
# hyperbolic
self.manifold = Hyperboloid()
# modules
self.emb_all = nn.Embedding(self.n_user, self.dim, padding_idx=0) # user embeddings
init_weight = self.init_size * torch.randn((self.n_user, self.dim))
# init_weight = self.manifold.expmap0(self.manifold.proj_tan0(init_weight, self.c), self.c)
self.emb_all.weight.data = init_weight
self.rot_src = nn.Embedding(self.n_user, self.dim) # user rotation matrices
self.rot_src.weight.data[:, ::2] = 1.0
self.rot_src.weight.data[:, 1::2] = 0.0
self.rot_tgt = nn.Embedding(self.n_user, self.dim)
self.rot_tgt.weight.data[:, ::2] = 1.0
self.rot_tgt.weight.data[:, 1::2] = 0.0
self.hyp_emb_module = HypUserEmbedding(
c=self.c, user_emb=self.emb_all, rot_matrix=[self.rot_src, self.rot_tgt],
n_neg=config['n_neg'], device=self.device)
# Attention Module
self.hyp_att = LorentzSelfAttention(c=self.c, dimension=self.dim)
self.time_att = TimeModule(emb_module=self.emb_all, att_layer=config['n_time_att_layer'],
att_heads=config['n_time_att_heads'], dropout=config['time_drop_out'],
device=self.device)
# rotary positional encoding
self.pos_encode = torch.randn(self.rotary_position_steps, self.dim, device=self.device)
self._set_angles(step=self.rotary_position_steps)
self.dropout_layer = nn.Dropout(self.dropout) # dropout layer
# Hawkes process
self.influence_intensity_hidden = nn.Linear(self.dim, self.n_user)
self.temporal_intensity_hidden = nn.Linear(self.dim, self.n_user)
self.factor_intensity_decay = torch.empty([1, self.n_user], device=self.device)
nn.init.xavier_normal_(self.factor_intensity_decay)
self.softplus = nn.Softplus()
# predictions
# CAT MANNER
# self.decode_user = nn.Linear(2*self.dim, self.n_user)
# PLUST MANNER
self.decode_user = nn.Linear(self.dim, self.n_user)
self.decode_time_user = nn.Linear(self.dim, self.n_user)
# time prediction
self.decode_time = nn.Linear(self.dim, 1)
# score weights
self.w_A_user = config['w_A_user']
# weight initialization
# self.decode_user.weight.data.fill_(0.01)
# torch.nn.init.xavier_normal_(self.decode_user.weight)
# torch.nn.init.xavier_uniform_(self.decode_user.weight)
def UserPrediction(self, hid_embeds, time_hiddens):
# dis_reg = distance_prediction_scores(hid_embeds, self.emb_all.weight)
pred_user_1 = self.decode_user(hid_embeds)
pred_user_2 = self.decode_time_user(time_hiddens)
pred_user = self.w_A_user * pred_user_1 + (1 - self.w_A_user) * pred_user_2
# pred_user = dis_reg + pred_user
# a = 0.95
# pred_user = a * pred_user_1 + (1-a) * pred_user_2
# pred_user = self.decode_user(hid_embeds + time_hiddens)
# hiddens = torch.cat([hid_embeds, time_hiddens], dim=-1)
# pred_user = self.decode_user(hiddens)
return pred_user
def TimePrediction(self, hid_embeds):
pred_time = self.decode_time(hid_embeds)
return pred_time
def forward(self, src_cas, src_time, tgt_time, mode='train'):
# data process
delta_times = tgt_time - src_time
delta_times = delta_times / (tgt_time + 1e-6)
# masks
att_mask = (src_cas == constants.PAD)
pred_mask = get_previous_user_mask(src_cas, self.n_user)
# map users in cascades to hyperbolic user representations
user_embeds_in_cas = self.emb_all(src_cas)
# calculate hidden representations (influence, temporal)
pos_user_embeds_in_cas = self.RotaryPositionalEncoding(user_embeds_in_cas)
inf_hiddens = self.hyp_att(pos_user_embeds_in_cas, pos_user_embeds_in_cas, pos_user_embeds_in_cas, att_mask)
time_hiddens = self.time_att(src_cas, src_time)
# next user prediction
pred_user = self.UserPrediction(inf_hiddens, time_hiddens)
pred_user = pred_user + pred_mask
# next time prediction
pred_time = self.TimePrediction(inf_hiddens)
if mode == 'train':
# use Hawkes process to fit observed historical cascades in training phrase
valid_mask = ~att_mask
type_mask = make_type_mask_for_pad_sequence(src_cas, self.n_user)
hawkes_loss = self.LogLikelihoodLoss(inf_states=inf_hiddens, time_states=time_hiddens,
time_delta_seqs=delta_times, pad_mask=valid_mask, type_mask=type_mask)
# hawkes_loss = None
return pred_user, pred_time, hawkes_loss
else:
return pred_user, pred_time
def LogLikelihoodLoss(self, inf_states, time_states, time_delta_seqs, pad_mask, type_mask):
""" calculate the loss for cascade fitting via the hyperbolic hawkes process"""
# prefer_states = 10 / prefer_states
factor_intensity_decay = self.factor_intensity_decay[None, ...]
intensity_states = factor_intensity_decay * time_delta_seqs[..., None] + self.temporal_intensity_hidden(
time_states) + self.influence_intensity_hidden(inf_states)
lambda_at_event = self.softplus(intensity_states)
# MCMC to estimate the Lambda
sample_dtimes = self.make_dtime_loss_samples(time_delta_seqs)
state_t_sample = self.compute_states_at_sample_times(
inf_states=inf_states, event_states=time_states, sample_dtimes=sample_dtimes)
lambda_t_sample = self.softplus(state_t_sample)
event_ll, non_event_ll, _ = compute_loglikelihood(
lambda_at_event=lambda_at_event, lambdas_loss_samples=lambda_t_sample, time_delta_seq=time_delta_seqs,
seq_mask=pad_mask, lambda_type_mask=type_mask)
loss = - (event_ll - non_event_ll).sum()
return loss
def make_dtime_loss_samples(self, time_delta_seq):
"""Generate the time point samples for every interval. """
dtimes_ratio_sampled = torch.linspace(start=0.0, end=1.0, steps=self.loss_integral_num_sample_per_step,
device=self.device)[None, None, :] # [1, 1, n_samples]
sampled_dtimes = time_delta_seq[:, :, None] * dtimes_ratio_sampled
return sampled_dtimes # [batch_size, max_len, n_samples]
def compute_states_at_sample_times(self, inf_states, event_states, sample_dtimes):
"""Compute the hidden states at sampled times. """
event_states = event_states[:, :, None, :] # [batch_size, seq_len, 1, hidden_size]
inf_states = inf_states[:, :, None, :]
sample_dtimes = sample_dtimes[..., None] # [batch_size, seq_len, num_samples, 1]
factor_intensity_decay = self.factor_intensity_decay[None, None, ...] # [1, 1, 1, num_event_types]
# [batch_size, seq_len, num_samples, num_event_types]
intensity_states = factor_intensity_decay * sample_dtimes + self.temporal_intensity_hidden(
event_states) + self.influence_intensity_hidden(inf_states)
return intensity_states
def train_emb(self, graph):
graph = graph.to(self.device)
emb_loss = self.hyp_emb_module(graph)
return emb_loss
def RotaryPositionalEncoding(self, x):
B, L, dim = x.size()
seq_idx = torch.arange(L).expand(B, L).to(x.device)
pos_rot_mat = F.embedding(seq_idx, self.pos_encode)
rot_pos_emb, x = pos_rot_mat.view(-1, self.dim), x.view(-1, self.dim)
pos_user_emb = givens_rotations(rot_pos_emb, x).reshape(B, L, -1)
return pos_user_emb
def _set_angles(self, step=20):
num = self.pos_encode.size(0)
repeat = int(num / step)
base_angle = math.pi / step
angles = torch.arange(step) * base_angle
angles = angles.unsqueeze(0).expand(repeat, step).reshape(num, 1)
msin = torch.sin(angles)
mcos = torch.cos(angles)
self.pos_encode[:, ::2] = mcos
self.pos_encode[:, 1::2] = msin