-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain_SimpleHHEA.py
158 lines (137 loc) · 6.49 KB
/
main_SimpleHHEA.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
import os
import random
import argparse
import numpy as np
from tqdm import tqdm
import torch
from CSLS_ import eval_alignment_by_sim_mat
from model import Simple_HHEA
from utils import *
### load embeddings
def noise_name_emb(name_emb, noise_ratio, emb_size=64):
sample_list = [i for i in range(emb_size)]
mask_id = random.sample(sample_list, int(emb_size * noise_ratio))
name_emb[:, mask_id] = 0
return name_emb
def load_embeddings(data_path, add_noise, noise_ratio, use_structure=True, use_time=True):
ent_name_emb, ent_dw_emb, ent_time_emb = None, None, None
### load name embeddings
kg1_name_emb = np.loadtxt(os.path.join(data_path, "ent_1_emb_64.txt"))
kg2_name_emb = np.loadtxt(os.path.join(data_path, "ent_2_emb_64.txt"))
ent_name_emb = np.array(kg1_name_emb.tolist() + kg2_name_emb.tolist())
print(f"read entity name embedding shape: {ent_name_emb.shape}")
if add_noise:
ent_name_emb = noise_name_emb(ent_name_emb, noise_ratio)
### load structure embeddings
if use_structure:
ent_dw_emb = np.loadtxt(os.path.join(data_path, "deep_emb.txt"))
print(f"read entity deepwalk emb shape: {ent_dw_emb.shape}")
### load time embeddings
if use_time:
ent_time_emb = np.array(load_ent_time_matrix(data_path))
print(f"read entity time embedding shape: {ent_time_emb.shape}")
return ent_name_emb, ent_dw_emb, ent_time_emb
### training
def l1(ll, rr):
return torch.sum(torch.abs(ll - rr), axis=-1)
def evaluate(model, dev_alignments, hit_k=[1, 5, 10], num_threads=16, csls=10):
model.eval()
with torch.no_grad():
feat = model()[dev_alignments]
Lvec, Rvec = feat[:, 0, :].detach().cpu().numpy(), feat[:, 1, :].detach().cpu().numpy()
Lvec = Lvec / np.linalg.norm(Lvec, axis=-1, keepdims=True)
Rvec = Rvec / np.linalg.norm(Rvec, axis=-1, keepdims=True)
t_prec_set, acc, t_mrr = eval_alignment_by_sim_mat(Lvec, Rvec, hit_k, num_threads, csls, accurate=True)
acc = [round(n, 3) for n in acc]
return t_prec_set, acc, t_mrr
def train(model:nn.Module, alignment_pairs, dev_alignments, epochs=1500, learning_rate=0.01, weight_decay=0.001, gamma=1.0, hit_k=[1, 5, 10]):
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
print(f"parameters: {get_n_params(model)}")
losses = []
t_prec = []
accs = []
t_mrrs = []
best_acc = [0] * len(hit_k)
best_mrr = 0
batch_size = len(alignment_pairs)
for i in tqdm(range(epochs)):
### forwad
model.train()
optimizer.zero_grad()
feat = model()[alignment_pairs]
### loss
l, r, fl, fr = feat[:, 0, :], feat[:, 1, :], feat[:, 2, :], feat[:, 3, :]
loss = torch.sum(nn.ReLU()(gamma + l1(l, r) - l1(l, fr)) + nn.ReLU()(gamma + l1(l, r) - l1(fl, r))) / batch_size
### backward
losses.append(loss.item())
loss.backward(retain_graph=True)
optimizer.step()
### evaluate
if (i + 1) % 10 == 0:
t_prec_set, acc, t_mrr = evaluate(model, dev_alignments, hit_k)
for i in range(len(hit_k)):
if best_acc[i] < acc[i]:
best_acc[i] = acc[i]
if best_mrr < t_mrr:
best_mrr = t_mrr
print(f"//best results: hits@{hit_k} = {best_acc}, mrr = {best_mrr:.3f}//")
accs.append(acc)
t_mrrs.append(t_mrr)
t_prec.append(t_prec_set)
return losses, t_prec, accs, t_mrrs, best_acc, best_mrr
if __name__ == "__main__":
### hyper parmeters
parser = argparse.ArgumentParser(description="Simple-HHEA Experiment")
parser.add_argument("--data", type=str, default="icews_wiki")
parser.add_argument("--cuda", type=int, default=3)
parser.add_argument("--random_seed", type=int, default=12306)
###### ablation settings
parser.add_argument("--add_noise", action="store_true")
parser.add_argument("--noise_ratio", type=float, default=0.3)
parser.add_argument("--no_structure", action="store_true")
parser.add_argument("--no_time", action="store_true")
###### training settings
parser.add_argument("--lr", type=float, default=0.01)
parser.add_argument("--wd", type=float, default=0.001)
parser.add_argument("--gamma", type=float, default=1.0)
parser.add_argument("--epochs", type=int, default=1500)
args = parser.parse_args()
### basic settings
data = args.data
device = f"cuda:{args.cuda}" if torch.cuda.is_available() and args.cuda >= 0 else "cpu"
use_time = ("icews_wiki" in data or "icews_yago" in data) and not args.no_time
use_structure = not args.no_structure
print(f"start exp: noise_ratio={args.noise_ratio}, data=\"{args.data}\", use_structure={use_structure}, use_time={use_time}")
### random settings
fixed(args.random_seed)
### load datas
data_path = os.path.join("data", data)
all_triples, node_size, rel_size = load_triples(data_path, True)
print(f"node_size={node_size} , rel_size={rel_size}")
train_alignments = load_alignments(os.path.join(data_path, "sup_pairs"))
dev_alignments = load_alignments(os.path.join(data_path, "ref_pairs"))
print(f"Train/Val: {len(train_alignments)}/{len(dev_alignments)}")
### load name embeddings
ent_name_emb, ent_dw_emb, ent_time_emb = load_embeddings(data_path, args.add_noise, args.noise_ratio, use_structure, use_time)
### model
model = Simple_HHEA(
time_span=1+27*13,
ent_name_emb=ent_name_emb,
ent_time_emb=ent_time_emb,
ent_dw_emb=ent_dw_emb,
use_structure=use_structure,
use_time=use_time,
emb_size=64,
structure_size=8,
time_size=8,
device=device
)
model = model.to(device)
alignment_pairs = get_train_set(train_alignments, node_size, node_size)
losses, t_prec, accs, t_mrrs, best_acc, best_mrr = train(model, alignment_pairs, dev_alignments, args.epochs, args.lr, args.wd, args.gamma, hit_k=[1, 5, 10])
### save result
result_dir = "result"
if not os.path.exists(result_dir):
os.makedirs(result_dir)
with open(os.path.join(result_dir, f"{data}_result_file_mlp.txt"), "a+", encoding="utf-8") as fw:
fw.write(f"settings: noise_ratio: {args.noise_ratio}, use_time: {use_time}, use_structure: {use_structure}\n\tbest results: hits@[1, 5, 10] = {best_acc}, mrr = {best_mrr:.3f}\n")