-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
186 lines (151 loc) · 6.8 KB
/
trainer.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
from cmath import log
import os
import time
import numpy as np
import random
import torch.nn.functional as F
import torch
from torch.utils.data import DataLoader
import torch.nn as nn
from transformers import AutoTokenizer
from dataset.reasoning_HR_mask_dataset import merge_data, ReasoningHRDataset
from model.gcplm import GCPLM
from utiles.args_utiles import get_args
from utiles.utiles_tools import load_triples
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
setup_seed(1024)
class Trainer:
def __init__(self, train_dataset, valid_dataset, args, test_dataset=None):
self.train_dataset = train_dataset
self.valid_dataset = valid_dataset
self.test_dataset = test_dataset
self.args = args
self.model_path = args.model_path
self.after_best_step = 0
self.train_load = DataLoader(self.train_dataset, batch_size=args.batch_size, collate_fn=merge_data,shuffle=True)
self.valid_load = DataLoader(self.valid_dataset, batch_size=args.batch_size, collate_fn=merge_data,shuffle=True)
self.valid_best_perform = -1
self.save_path = str(os.path.join(self.args.ckpt_saving_path,"{}_checkpoint_{}_{}_{}.reasoning".format(self.args.lang, self.args.e_measure,self.args.p_model, self.args.tuning_mode)))
if not os.path.exists(self.save_path):
os.mkdir(self.save_path)
self.model_initial()
self.log_init()
def log_init(self):
with open(os.path.join(self.save_path,"log"),"w") as f:
f.write("creat file \n")
def model_initial(self):
self.model = GCPLM(args, self.model_path)
self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=args.lr)
if self.args.use_cuda:
if self.args.gpus:
self.model = nn.DataParallel(self.model)
self.model = self.model.cuda()
else:
self.model = self.model.cuda()
def save_model(self, current_perform):
if current_perform > self.valid_best_perform:
self.after_best_step = 0
self.valid_best_perform = current_perform
self.log("Saving best:\t {}\n".format(self.valid_best_perform))
self.model.model.save_pretrained(self.save_path)
else:
self.after_best_step += 1
def log(self,log_content):
with open(os.path.join(self.save_path,"log"),"a") as f:
f.write(str(log_content)+"\n")
print(str(log_content))
def tensor_to_cuda(self,inpt,inpt_info):
'''
{
"input_ids": batch_inpt_ids,
"attention_mask": batch_mask_attention,
"batch_label":torch.LongTensor(batch_label),
"batch_label_index":torch.LongTensor(batch_label_index),
"batch_reasoning_index":torch.LongTensor(batch_reasoing_index),
"batch_tail_index": torch.LongTensor(batch_tail_index[:-1])
}
'''
for k in inpt.keys():
inpt[k] = inpt[k].cuda()
for k in inpt_info.keys():
inpt_info[k] = inpt_info[k].cuda()
return inpt, inpt_info
def train(self):
self.model.train()
self.optimizer.zero_grad()
num_step = 0
for epoch in range(1, self.args.epoch):
train_epoch_losses = 0
loss_iterm_dic = {"taill loss":0, "reasoning loss":0, "global local loss":0}
for batch in self.train_load:
inpt, inpt_info = batch
if args.use_cuda:
batch = self.tensor_to_cuda(inpt, inpt_info)
loss,loss_iterm = self.model(inpt, inpt_info)
train_epoch_losses += loss.cpu().item() / len(self.train_load)
for k in loss_iterm:
loss_iterm_dic[k] += loss_iterm[k].cpu().item()/len(self.train_load)
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
self.log('**{}*******************{}************************'.format(epoch,time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())))
self.log("epoch:{}, whole loss : {}".format(epoch, train_epoch_losses))
for kw in loss_iterm_dic:
self.log("epoch:{}, {} : {}".format(epoch, kw,loss_iterm_dic[kw]))
# wandb.log({"tail_loss": train_epoch_tail_losses})
self.eval(epoch)
if self.after_best_step > self.args.early_stop:
self.log("Early stopping is triggered in epoch:{}, after best step {}".format(str(epoch),str(self.after_best_step)))
break
def eval(self,epoch=-1):
self.model.eval()
pred_acc = []
score_list = []
for batch in self.valid_load:
inpt, inpt_info = batch
if args.use_cuda:
batch = self.tensor_to_cuda(inpt, inpt_info)
out, score_a = self.model(inpt, inpt_info, mod="eval")
pred_t = torch.argmax(out, dim=-1)
a = pred_t == inpt_info["batch_label"]
score = score_a.norm(p=1,dim=0).mean(dim=0)
pred_acc.append(torch.sum(a).item() / len(a))
score_list.append(score.item())
acc = sum(pred_acc) / len(pred_acc)
# wandb.log({"acc": acc})
self.log("epoch {}, ACC:{}".format(epoch,acc))
self.log('epoch {}, SCORE{}'.format(epoch,sum(score_list) / len(score_list)))
self.save_model(acc)
def test(self):
pass
if __name__ == '__main__':
args = get_args()
# wandb.init(project="prix_potint_reasoing_addtoks_{}".format(args.lang), entity="maxpain")
train_triples = load_triples(args.raw_train_path, args.lang)
valid_triples = load_triples(args.raw_valid_path, args.lang)
# test_triples = load_triples(args.raw_test_path, args.lang)
# train_triples, valid_triples = train_triples[200:], train_triples[:200]
e1_mark = ['[S]']
r_mark = ['[P]']
e2_mark = ['[O]']
eos_mark = ['[EOS]']
tokenizer = AutoTokenizer.from_pretrained(args.model_path)
for i in ["[S]","[P]","[O]","[EOS]"]:
tokenizer.add_tokens(i)
args.embedding_size = len(tokenizer)
# 没有 HR mask
# triple_dataset = ReasoningDataset(train_triples, tokenizer, args)
# valid_dataset = ReasoningDataset(valid_triples, tokenizer, args) ReasoningHRDataset
triple_dataset = ReasoningHRDataset(train_triples, tokenizer, args)
valid_dataset = ReasoningHRDataset(valid_triples, tokenizer, args)
# new_tokens_sorted = sorted(new_tokens.items(), key=lambda x: x[1], reverse=False)
# print(new_tokens_sorted)
# tokenizer.add_tokens([i[0] for i in new_tokens_sorted])
trainer = Trainer(triple_dataset, valid_dataset, args)
trainer.train()
trainer.eval()