-
Notifications
You must be signed in to change notification settings - Fork 0
/
finetune_T5.py
212 lines (166 loc) · 8.97 KB
/
finetune_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
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
206
207
208
209
210
211
212
from transformers import T5Tokenizer, T5ForConditionalGeneration
import transformers
from transformers import Adafactor
import os
import json
import torch
import torch.optim as optim
import pickle
import torch.nn as nn
import random
import wandb
from torch.utils.data import Dataset
import argparse
from torch.utils.data import DataLoader
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('--exp_name',type=str,default=None)
parser.add_argument('--eval_bs', type = int, default = 3)
parser.add_argument('--bs', type = int, default = 5)
parser.add_argument('--backpropagate', type = int, default = 5)
#parser.add_argument('--store',type=str,default=None)
parser.add_argument("--print_every", type = int, default=1000)
parser.add_argument('--eval_every',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)
class Data(Dataset):
def __init__(self, data):
self.data = data
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx][0], self.data[idx][1]
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([data['text'], data['summary']])
return lis
data_train = Data(read_data(args.train_file))
data_dev = read_data(args.dev_file)[:1000]
# for i in range(len(data_train)):
# data_train[i][0] = data_train[i][0].replace('<extra_id_','T').replace('>','')
# data_train[i][1] = data_train[i][1].replace('<extra_id_','T').replace('>','')
# for i in range(len(data_dev)):
# data_dev[i][0] = data_dev[i][0].replace('<extra_id_','T').replace('>','')
# data_dev[i][1] = data_dev[i][1].replace('<extra_id_','T').replace('>','')
class Model(nn.Module):
def __init__(self,model_name):
super(Model,self).__init__()
self.model=T5ForConditionalGeneration.from_pretrained(model_name)
def forward(self,input):
outputs=self.model(input_ids=input['input_ids'], \
labels=input['labels'], attention_mask=input['attention_mask'],output_hidden_states=True,output_attentions=True)
return outputs.loss
class Train:
def __init__(self,data,data_val,args):
self.data=data
self.dev_data=data_val
self.args=args
self.tokenizer=T5Tokenizer.from_pretrained(args.model_name)
self.model=nn.DataParallel(Model(args.model_name),device_ids=[args.device])
self.model.to(f'cuda:{self.model.device_ids[0]}')
self.optimizer=optim.AdamW(self.model.parameters(),lr=0.0015) #0.00015 for cwq
self.lr_scheduler=transformers. \
get_polynomial_decay_schedule_with_warmup(self.optimizer, 5000, 30000,power=0.5) # 30000 for all except lcq1
'''self.optimizer=Adafactor(self.model.parameters(),lr=1e-2,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.epochs = 100
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.backpropagate=args.backpropagate
self.train_dataloader = DataLoader(data, batch_size=self.bs, shuffle=True)
self.train()
def generate_batch(self):
output=random.sample(self.data,self.bs)
inp,label=[],[]
for dat in output:
inp.append(dat[0])
label.append(dat[1]) #.replace('<extra_id_','T').replace('>',''))
return inp,label
def preprocess_function(self,inputs, targets):
model_inputs=self.tokenizer(inputs, padding=True, \
return_tensors='pt',max_length=512, truncation=True)
labels=self.tokenizer(targets,padding=True,max_length=512, truncation=True)
if True:
labels["input_ids"] = [
[(l if l != self.tokenizer.pad_token_id else -100) \
for l in label] for label in labels["input_ids"]
]
labels['input_ids']=torch.tensor(labels['input_ids'])
model_inputs["labels"]=labels["input_ids"].to(f'cuda:{self.model.device_ids[0]}')
model_inputs["input_ids"]=model_inputs["input_ids"].to(f'cuda:{self.model.device_ids[0]}')
model_inputs["attention_mask"]=model_inputs["attention_mask"].to(f'cuda:{self.model.device_ids[0]}')
return model_inputs
def val(self,o):
self.model.eval()
acc,bs,i=0,self.eval_bs,0
saver=[]
while i<len(self.dev_data):
bs_=min(bs,len(self.dev_data)-i)
i+=bs_
inp,label=[],[]
for j in range(i-bs_,i):
inp.append(self.dev_data[j][0])
label.append(self.dev_data[j][1]) #.replace('<extra_id_','T').replace('>',''))
input=self.preprocess_function(inp,label)
output=self.model.module.model.generate(input_ids=input['input_ids'],
num_beams=10,attention_mask=input['attention_mask'], \
early_stopping=True, max_length=400,output_hidden_states=True,output_attentions=True)
out=self.tokenizer.batch_decode(output,skip_special_tokens=False)
for k in range(len(out)):
#print(out[k].replace('<pad>','').replace('</s>','').strip())
a1=out[k].replace('<pad>','').replace('</s>','').replace('<unk>','').replace('<s>','').strip().replace(' ','')
a2=label[k].strip().replace(' ','')
#print(a1, ' ', a2)
saver.append({'input':inp[k],'gold':label[k].strip(),'generated':out[k].replace('<pad>',''). \
replace('</s>','').replace('<unk>','').replace('<s>','').strip()})
if a1==a2:
acc+=1; #print('ttt')
file=open(self.args.exp_name+'/'+str(o)+'dev_result.json','w')
json.dump(saver,file)
file.close()
wandb.log({"epochs": o, "matches": acc})
return acc
def train(self):
loss, j, tot_loss, log_loss, prev_acc = 0, 0, 0, 0, 0
for epoch in range(self.epochs):
for inp, label in self.train_dataloader:
self.model.train()
input = self.preprocess_function(inp,label)
loss_temp = self.model(input)
tot_loss += loss_temp.item()
loss += loss_temp/self.backpropagate
if(j+1)%self.print_every==0:
print('epoch = {}, iteration = {}, training loss = {}'.format(epoch, j, (tot_loss-log_loss)/self.print_every))
log_loss = tot_loss
if (j+1)%self.backpropagate == 0:
loss.backward()
self.optimizer.step()
self.lr_scheduler.step()
self.optimizer.zero_grad()
loss = 0
j+=1
if(epoch+1)%self.eval_every==0:
acc=self.val(epoch)
print('validation acc={}'.format(acc))
if prev_acc <= acc:
prev_acc = acc
torch.save(self.model.state_dict(),self.args.exp_name+'/'+'checkpoint.pth')
trainer=Train(data_train,data_dev,args)