-
Notifications
You must be signed in to change notification settings - Fork 4
/
train_other_tokenizer.py
205 lines (183 loc) · 8.46 KB
/
train_other_tokenizer.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
from dataset import VTKG
from model_mygo import MyGO
from tqdm import tqdm
from utils import calculate_rank, metrics
import numpy as np
import argparse
import torch
import torch.nn as nn
import datetime
import time
import os
import copy
import math
import random
import distutils
import logging
from merge_tokens import get_entity_visual_tokens, get_entity_textual_tokens
OMP_NUM_THREADS=8
torch.backends.cudnn.benchmark = True
torch.set_num_threads(8)
torch.cuda.empty_cache()
torch.manual_seed(2024)
random.seed(2024)
np.random.seed(2024)
logger = logging.getLogger()
logger.setLevel(logging.INFO)
log_format = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s')
stream_handler = logging.StreamHandler()
stream_handler.setFormatter(log_format)
logger.addHandler(stream_handler)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--data', default="MKG-W", type = str)
parser.add_argument('--lr', default=5e-4, type=float)
parser.add_argument('--dim', default=200, type=int)
parser.add_argument('--num_epoch', default=100, type=int)
parser.add_argument('--valid_epoch', default=50, type=int)
parser.add_argument('--exp', default='mygo')
parser.add_argument('--no_write', action='store_true')
parser.add_argument('--num_layer_enc_ent', default=1, type=int)
parser.add_argument('--num_layer_enc_rel', default=1, type=int)
parser.add_argument('--num_layer_dec', default=2, type=int)
parser.add_argument('--num_head', default=2, type=int)
parser.add_argument('--hidden_dim', default=200, type = int)
parser.add_argument('--dropout', default=0.01, type = float)
parser.add_argument('--emb_dropout', default=0.9, type = float)
parser.add_argument('--vis_dropout', default=0.4, type = float)
parser.add_argument('--txt_dropout', default=0.1, type = float)
parser.add_argument('--smoothing', default=0.0, type = float)
parser.add_argument('--batch_size', default=2048, type = int)
parser.add_argument('--decay', default=0.0, type = float)
parser.add_argument('--max_img_num', default=3, type = int)
parser.add_argument('--cont', action = 'store_true')
parser.add_argument('--step_size', default=50, type = int)
parser.add_argument('--max_vis_token', default=8, type=int)
parser.add_argument('--max_txt_token', default=8, type=int)
parser.add_argument('--score_function', default="tucker", type=str)
parser.add_argument('--mu', default=0, type=float)
parser.add_argument('--text_tokenizer', default="bert", type=str)
parser.add_argument('--visual_tokenizer', default="beit", type=str)
args = parser.parse_args()
file_format = ""
for arg_name in vars(args).keys():
if arg_name in ["lr", "hidden_dim", "batch_size", "num_epoch", "max_vis_token", "max_txt_token", "num_head", "mu"]:
file_format += f"{arg_name}_{vars(args)[arg_name]}"
elif arg_name in ["score_function", "emb_dropout", "vis_dropout", "txt_dropout"]:
file_format += f"{vars(args)[arg_name]}"
if not args.no_write:
os.makedirs(f"./result/{args.exp}/{args.data}", exist_ok = True)
os.makedirs(f"./ckpt/{args.exp}/{args.data}", exist_ok = True)
os.makedirs(f"./logs/{args.exp}/{args.data}", exist_ok = True)
if not os.path.isfile(f"ckpt/{args.exp}/args.txt"):
with open(f"ckpt/{args.exp}/args.txt", "w") as f:
for arg_name in vars(args).keys():
if arg_name not in ["data", "exp", "no_write", "num_epoch", "cont", "early_stop"]:
f.write(f"{arg_name}\t{type(vars(args)[arg_name])}\n")
else:
file_format = None
file_handler = logging.FileHandler(f"./logs/{args.exp}/{args.data}/{file_format}.log")
file_handler.setFormatter(log_format)
logger.addHandler(file_handler)
logger.info(f"{os.getpid()}")
logger.info(args)
KG = VTKG(args.data, logger, max_vis_len = args.max_img_num)
KG_Loader = torch.utils.data.DataLoader(KG, batch_size = args.batch_size, shuffle=True)
visual_token_index, visual_key_mask = get_entity_visual_tokens(
dataset=args.data,
max_num=args.max_vis_token,
type=args.visual_tokenizer
)
visual_token_index = visual_token_index.cuda()
text_token_index, text_key_mask = get_entity_textual_tokens(
dataset=args.data,
max_num=args.max_txt_token,
type=args.text_tokenizer
)
text_token_index = text_token_index.cuda()
logger.info(visual_token_index, text_token_index)
logger.info(visual_key_mask, text_key_mask)
model = MyGO(
num_ent = KG.num_ent,
num_rel = KG.num_rel,
ent_vis_mask = visual_key_mask,
ent_txt_mask = text_key_mask,
dim_str = args.dim,
num_head = args.num_head,
dim_hid = args.hidden_dim,
num_layer_enc_ent = args.num_layer_enc_ent,
num_layer_enc_rel = args.num_layer_enc_rel,
num_layer_dec = args.num_layer_dec,
dropout = args.dropout,
emb_dropout = args.emb_dropout,
vis_dropout = args.vis_dropout,
txt_dropout = args.txt_dropout,
visual_token_index = visual_token_index,
text_token_index = text_token_index,
score_function = args.score_function,
text_tokenizer = args.text_tokenizer,
visual_tokenizer= args.visual_tokenizer
).cuda()
logger.info(model)
loss_fn = nn.CrossEntropyLoss(label_smoothing = args.smoothing)
optimizer = torch.optim.Adam(model.parameters(), lr = args.lr, weight_decay = args.decay)
scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, args.step_size, T_mult = 2)
last_epoch = 0
start = time.time()
logger.info("EPOCH\tLOSS\tTOTAL TIME")
all_ents = torch.arange(KG.num_ent).cuda()
all_rels = torch.arange(KG.num_rel).cuda()
best_mrr = 0.0
for epoch in range(last_epoch + 1, args.num_epoch + 1):
total_loss = 0.0
for batch, label in KG_Loader:
ent_embs, rel_embs = model()
scores = model.score(ent_embs, rel_embs, batch.cuda())
loss = loss_fn(scores, label.cuda())
if args.mu != 0:
loss_con = model.contrastive_loss_finegrained(ent_embs)
loss += loss_con * args.mu
total_loss += loss.item()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
scheduler.step()
logger.info(f"{epoch} \t {total_loss:.6f} \t {time.time() - start:.6f} s")
if (epoch) % args.valid_epoch == 0:
model.eval()
with torch.no_grad():
ent_embs, rel_embs = model()
lp_list_rank = []
for triplet in tqdm(KG.test):
h,r,t = triplet
head_score = model.score(ent_embs, rel_embs, torch.tensor([[KG.num_ent + KG.num_rel, r + KG.num_ent, t + KG.num_rel]]).cuda())[0].detach().cpu().numpy()
head_rank = calculate_rank(head_score, h, KG.filter_dict[(-1, r, t)])
tail_score = model.score(ent_embs, rel_embs, torch.tensor([[h + KG.num_rel, r + KG.num_ent, KG.num_ent + KG.num_rel]]).cuda())[0].detach().cpu().numpy()
tail_rank = calculate_rank(tail_score, t, KG.filter_dict[(h, r, -1)])
lp_list_rank.append(head_rank)
lp_list_rank.append(tail_rank)
lp_list_rank = np.array(lp_list_rank)
mr, mrr, hit10, hit3, hit1 = metrics(lp_list_rank)
logger.info("Link Prediction on Validation Set")
logger.info(f"MR: {mr}")
logger.info(f"MRR: {mrr}")
logger.info(f"Hit10: {hit10}")
logger.info(f"Hit3: {hit3}")
logger.info(f"Hit1: {hit1}")
if best_mrr < mrr:
best_mrr = mrr
best_result = (mr, mrr, hit10, hit3, hit1)
model.train()
if (epoch) % 500 == 0:
torch.save(
{
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict()
},
f"./ckpt/{args.exp}/{args.data}/{file_format}_{epoch}.ckpt"
)
model.train()
logger.info("Done! {}. The best results are shown below:".format(args.data))
logger.info(best_result)