/
Train_T5.py
180 lines (146 loc) · 7.98 KB
/
Train_T5.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
from transformers import T5Tokenizer, T5ForConditionalGeneration
import transformers
import os
import json
from transformers import Adafactor
import torch
import torch.optim as optim
import pickle
import torch.nn as nn
import random
from openprompt.plms import load_plm
from openprompt import PromptDataLoader
from openprompt.prompts.prefix_tuning_template import PrefixTuningTemplate
from openprompt import PromptForGeneration
from openprompt.data_utils.utils import InputExample
import argparse
import wandb
parser=argparse.ArgumentParser()
parser.add_argument('--train_file',type=str,default=None)
parser.add_argument('--dev_file',type=str,default=None)
parser.add_argument('--test',type=bool,default=False)
parser.add_argument('--model_name',type=str,default='t5-base')
parser.add_argument('--checkpoint',type=str,default=None)
parser.add_argument('--device',type=int,default=0)
parser.add_argument('--num_token',type=int,default=512)
parser.add_argument("--plm_eval_mode", action="store_true")
parser.add_argument('--exp_name',type=str,default=None)
parser.add_argument("--print_every", type = int, default=1000)
parser.add_argument('--eval_every',type=int,default=5)
parser.add_argument("--eval_bs", type = int, default=6)
parser.add_argument('--bs',type=int,default=5)
args=parser.parse_args()
#wandb.init(project='replacements_new_small_cwq',name=args.exp_name)
if os.path.exists(args.exp_name)==False:
os.mkdir(args.exp_name)
torch.manual_seed(42)
def read_data(filename):
f = open(filename)
lines = f.readlines()
lis = []
for i,item in enumerate(lines):
data = json.loads(item)
if not data['text'] or not data['summary']:
continue
lis.append(InputExample(guid = str(i),text_a = data['text'],tgt_text = data['summary']))
return lis
dataset={}
dataset['train'] = read_data(args.train_file)
dataset['validation'] = read_data(args.dev_file)[:1000]
class Train:
def __init__(self,dataset,args):
self.dataset = dataset
self.args=args
self.epochs = 200
self.print_every=args.print_every
self.eval_every=args.eval_every
self.num_gpus=1
self.eval_bs=args.eval_bs
self.bs=args.bs
self.back_propogate=2
self.use_cuda = True
plm, tokenizer, model_config, WrapperClass = load_plm(args.model_name.split('-')[0], args.model_name)
self.mytemplate = PrefixTuningTemplate(model=plm, num_token=args.num_token, tokenizer=tokenizer, placeholder_mapping = {'<text_a>': 'text_a', '<text_b>': 'text_b'})
self.train_dataloader = PromptDataLoader(dataset=dataset["train"], template=self.mytemplate, tokenizer=tokenizer,
tokenizer_wrapper_class=WrapperClass, max_seq_length=512, decoder_max_length=512,
batch_size=self.bs,shuffle=True, teacher_forcing=True, predict_eos_token=True, # be sure to pass predict_eos_token=True if your tempalte doesn't contain one, or you model may fail to stop generation.
truncate_method="head")
self.validation_dataloader = PromptDataLoader(dataset=dataset["validation"], template=self.mytemplate, tokenizer=tokenizer,
tokenizer_wrapper_class=WrapperClass, max_seq_length=512, decoder_max_length=512,
batch_size=self.eval_bs,shuffle=False, teacher_forcing=False, predict_eos_token=True,
truncate_method="head")
self.prompt_model = PromptForGeneration(plm=plm,template=self.mytemplate, freeze_plm=True,tokenizer=tokenizer, plm_eval_mode=args.plm_eval_mode)
if self.use_cuda:
self.prompt_model = self.prompt_model.cuda()
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in self.mytemplate.named_parameters() if (not any(nd in n for nd in no_decay)) and p.requires_grad],
"weight_decay": 0.0,
},
{
"params": [p for n, p in self.mytemplate.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad],
"weight_decay": 0.0,
},
]
#self.optimizer=optim.AdamW(optimizer_grouped_parameters, lr=0.00005)
self.optimizer=Adafactor(optimizer_grouped_parameters,lr=1e-3,eps=(1e-30, 1e-3),clip_threshold=1.0, \
beta1=0.0,weight_decay=0.0,relative_step=False, \
scale_parameter=True,warmup_init=False)
#self.lr_scheduler=transformers.get_polynomial_decay_schedule_with_warmup(self.optimizer, 5000, 30000, power = 0.5)
self.train()
def val(self,epoch):
generated_sentence = []
groundtruth_sentence = []
self.prompt_model.eval()
for step, inputs in enumerate(self.validation_dataloader):
if self.use_cuda:
inputs = inputs.cuda()
_,output_sentence=self.prompt_model.generate(inputs,
num_beams=10, \
early_stopping=True, max_length=512,output_hidden_states=True,output_attentions=True)
output_sentence=[o.replace('<unk>','').replace('<pad>','').replace('<s>','').replace('</s>','') for o in output_sentence]
gold = [ii.replace('<unk>','').replace('<pad>','').replace('<s>','').replace('</s>','') for ii in inputs['tgt_text']]
generated_sentence.extend(output_sentence)
groundtruth_sentence.extend(gold)
acc = 0
file=open(self.args.exp_name+'/'+str(epoch)+'gen.txt','w')
file1=open(self.args.exp_name+'/'+str(epoch)+'ref.txt','w')
for i in range(len(generated_sentence)):
file1.write(groundtruth_sentence[i].strip()+'\n')
file.write(generated_sentence[i].strip()+'\n')
if groundtruth_sentence[i].strip().replace(' ','') == generated_sentence[i].strip().replace(' ',''): acc+=1
file.close()
file1.close()
wandb.log({"epoch": epoch, "matches": acc })
print(acc)
return acc
def train(self):
global_step = 0
tot_loss = 0
log_loss = 0
prev_acc = 0
for epoch in range(self.epochs):
self.prompt_model.train()
for step, inputs in enumerate(self.train_dataloader):
global_step +=1
if self.use_cuda:
inputs = inputs.cuda()
loss = self.prompt_model(inputs)
#loss += 0.1*(self.mytemplate.loss + self.mytemplate.loss1)
loss.backward()
tot_loss += loss.item()
torch.nn.utils.clip_grad_norm_(self.mytemplate.parameters(), 1.0)
self.optimizer.step()
#self.lr_scheduler.step()
self.optimizer.zero_grad()
if (1+global_step) %self.print_every ==0:
print("Epoch {}, global_step {} average loss: {}".format(epoch, global_step, (tot_loss-log_loss)/self.print_every), flush=True)
#print("Epoch {}, global_step {} average loss: {} lr: {}".format(epoch, global_step, (tot_loss-log_loss)/500, scheduler.get_last_lr()[0]), flush=True)
log_loss = tot_loss
if (1+epoch) % self.eval_every ==0:
acc = self.val(epoch)
if prev_acc <= acc:
prev_acc = acc
torch.save(self.mytemplate.state_dict(),self.args.exp_name+'/'+'checkpoint.pth')
trainer=Train(dataset,args)