-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
165 lines (142 loc) · 8.15 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
import dgl
import torch
import torch.nn as nn
class IntraHierarchyCommunication(nn.Module):
def __init__(self, embedding_dim, h_size, nary):
super(IntraHierarchyCommunication, self).__init__()
self.nary = nary
self.W_f = nn.Linear(embedding_dim, h_size, bias=False) # W_f -> [embedding_dim, h_size]
self.U_f = nn.Linear(nary * h_size, nary * h_size, bias=False)
self.b_f = nn.Parameter(torch.zeros(1, h_size))
self.W_iou = nn.Linear(embedding_dim, 3 * h_size, bias=False) # [W_i, W_u, W_o] -> [embedding_dim, 3 * h_size]
self.U_iou = nn.Linear(nary * h_size, 3 * h_size, bias=False)
self.b_iou = nn.Parameter(torch.zeros(1, 3 * h_size))
# Transformer encoder
encoder_layers = nn.TransformerEncoderLayer(h_size, 2, h_size, 0.4)
self.transformer_encoder = nn.TransformerEncoder(encoder_layers, 2)
def apply_node_func(self, nodes):
iou = nodes.data["iou"]
i, o, u = torch.chunk(iou, 3, 1)
i, o, u = torch.sigmoid(i), torch.sigmoid(o), torch.tanh(u)
c = i * u + nodes.data["c"] # [batch, h_size]
h = o * torch.tanh(c)
return {"h": h, "c": c}
def message_func(self, edges):
return {"h_child": edges.src["h"], "c_child": edges.src["c"], "type": edges.src["type"]}
def reduce_func(self, nodes):
Wx = torch.cat([self.W_f(nodes.data["x"]) for _ in range(self.nary)], dim=1)
b_f = torch.cat([self.b_f for _ in range(self.nary)], dim=1)
h_cat = nodes.mailbox["h_child"] # [batch, nary, h_size]
h_cat = h_cat.view(h_cat.size(0), -1) # [batch, nary * h_size]
f = torch.sigmoid(Wx + self.U_f(h_cat) + b_f)
h_cat_att = self.transformer_encoder(nodes.mailbox["h_child"])
h_cat_att = h_cat_att.view(h_cat_att.size(0), -1)
iou = self.W_iou(nodes.data["x"]) + self.U_iou(h_cat_att) + self.b_iou # [batch, 3 * h_size]
c = torch.sum(f.view(nodes.mailbox["c_child"].size()) * nodes.mailbox["c_child"], 1)
return {"c": c.view(c.size(0), -1), "iou": iou}
class InterHierarchyCommunication(nn.Module):
def __init__(self, h_size, nary):
super(InterHierarchyCommunication, self).__init__()
self.nary = nary
self.W_f = nn.Linear(h_size, h_size, bias=False) # W_f -> [embedding_dim, h_size]
self.U_f = nn.Linear(nary * h_size, nary * h_size, bias=False)
self.b_f = nn.Parameter(torch.zeros(1, h_size))
self.W_iou = nn.Linear(h_size, 3 * h_size, bias=False) # [W_i, W_u, W_o] -> [embedding_dim, 3 * h_size]
self.U_iou = nn.Linear(nary * h_size, 3 * h_size, bias=False)
self.b_iou = nn.Parameter(torch.zeros(1, 3 * h_size))
# Transformer encoder
encoder_layers = nn.TransformerEncoderLayer(h_size, 2, h_size, 0.4)
self.transformer_encoder = nn.TransformerEncoder(encoder_layers, 2)
def apply_node_func(self, nodes):
iou = nodes.data["iou"]
i, o, u = torch.chunk(iou, 3, 1)
i, o, u = torch.sigmoid(i), torch.sigmoid(o), torch.tanh(u)
c = i * u + nodes.data["c"] # [batch, h_size]
h = o * torch.tanh(c)
return {"h": h, "c": c}
def message_func(self, edges):
return {"h_child": edges.src["h"], "c_child": edges.src["c"], "type": edges.src["type"]}
def reduce_func(self, nodes):
Wx = torch.cat([self.W_f(nodes.data["x"]) for _ in range(self.nary)], dim=1)
b_f = torch.cat([self.b_f for _ in range(self.nary)], dim=1)
h_cat = nodes.mailbox["h_child"] # [batch, nary, h_size]
h_cat = h_cat.view(h_cat.size(0), -1) # [batch, nary * h_size]
f = torch.sigmoid(Wx + self.U_f(h_cat) + b_f)
h_cat_att = self.transformer_encoder(nodes.mailbox["h_child"])
h_cat_att = h_cat_att.view(h_cat_att.size(0), -1)
iou = self.W_iou(nodes.data["x"]) + self.U_iou(h_cat_att) + self.b_iou # [batch, 3 * h_size]
c = torch.sum(f.view(nodes.mailbox["c_child"].size()) * nodes.mailbox["c_child"], 1)
return {"c": c.view(c.size(0), -1), "iou": iou}
class TreeLSTM(nn.Module):
def __init__(self,
h_size=512, nary=5,
embed_dropout=0.2, model_dropout=0.4,
num_users=3000, user_embed_dim=128,
num_POIs=5000, POI_embed_dim=128,
num_cats=300, cat_embed_dim=32,
num_coos=1024, coo_embed_dim=64,
device='cuda'):
super(TreeLSTM, self).__init__()
self.device = device
self.h_size = h_size
self.nary = nary
# embedding layer
self.embedding_dim = user_embed_dim + POI_embed_dim + cat_embed_dim + coo_embed_dim
self.user_embedding = nn.Embedding(num_embeddings=num_users, embedding_dim=user_embed_dim)
self.POI_embedding = nn.Embedding(num_embeddings=num_POIs, embedding_dim=POI_embed_dim)
self.cat_embedding = nn.Embedding(num_embeddings=num_cats, embedding_dim=cat_embed_dim)
self.coo_embedding = nn.Embedding(num_embeddings=num_coos, embedding_dim=coo_embed_dim)
# positional encoding layer
self.time_pos_encoder = nn.Embedding(num_embeddings=96, embedding_dim=self.embedding_dim) # 24*4
# dropout layer
self.embed_dropout = nn.Dropout(embed_dropout)
self.model_dropout = nn.Dropout(model_dropout)
# LSTM cell
self.cell_IAC = IntraHierarchyCommunication(self.embedding_dim, h_size, nary)
self.cell_IRC = InterHierarchyCommunication(h_size, nary)
# decoder layer
self.decoder_POI = nn.Linear(h_size, num_POIs)
self.decoder_cat = nn.Linear(h_size, num_cats)
self.decoder_coo = nn.Linear(h_size, num_coos)
def forward(self, MT_input):
user_embedding = self.user_embedding(MT_input.features[:, 0].long() * MT_input.mask)
POI_embedding = self.POI_embedding(MT_input.features[:, 1].long() * MT_input.mask)
cat_embedding = self.cat_embedding(MT_input.features[:, 2].long() * MT_input.mask)
coo_embedding = self.coo_embedding(MT_input.features[:, 3].long() * MT_input.mask)
pe = self.time_pos_encoder(MT_input.time.long() * MT_input.mask)
concat_embedding = torch.cat((user_embedding, POI_embedding, cat_embedding, coo_embedding), dim=1)
concat_embedding = concat_embedding + pe * 0.5
g = MT_input.graph.to(self.device)
n = g.num_nodes()
g.ndata["iou"] = self.cell_IAC.W_iou(self.embed_dropout(concat_embedding)) * MT_input.mask.float().unsqueeze(-1)
g.ndata["x"] = self.embed_dropout(concat_embedding) * MT_input.mask.float().unsqueeze(-1)
g.ndata["h"] = torch.zeros((n, self.h_size)).to(self.device)
g.ndata["c"] = torch.zeros((n, self.h_size)).to(self.device)
g.ndata["h_child"] = torch.zeros((n, self.nary, self.h_size)).to(self.device)
g.ndata["c_child"] = torch.zeros((n, self.nary, self.h_size)).to(self.device)
dgl.prop_nodes_topo(graph=g,
message_func=self.cell_IAC.message_func,
reduce_func=self.cell_IAC.reduce_func,
apply_node_func=self.cell_IAC.apply_node_func)
h_1 = g.ndata["h"] # [batch_size, h_size]
g.ndata["x"] = h_1 * MT_input.mask2.float().unsqueeze(-1)
dgl.prop_nodes_topo(graph=g,
message_func=self.cell_IRC.message_func,
reduce_func=self.cell_IRC.reduce_func,
apply_node_func=self.cell_IRC.apply_node_func)
h_2 = self.model_dropout(g.ndata["h"]) # [batch_size, h_size]
y_pred_POI = self.decoder_POI(h_2)
y_pred_cat = self.decoder_cat(h_2)
y_pred_coo = self.decoder_coo(h_2)
return y_pred_POI, y_pred_cat, y_pred_coo
class MultiTaskLoss(nn.Module):
def __init__(self, num=3):
super(MultiTaskLoss, self).__init__()
params = torch.ones(num, requires_grad=True)
self.params = nn.Parameter(params)
def forward(self, *losses):
loss_sum = 0
for i, loss in enumerate(losses):
# loss_sum += 0.5 / (self.params[i] ** 2) * loss + torch.log(1 + self.params[i] ** 2)
loss_sum += 0.5 * torch.exp(-self.params[i]) * loss + self.params[i]
return loss_sum