-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
174 lines (141 loc) · 5.83 KB
/
main.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
import tqdm
import random
import yaml
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.optim
import torch.nn.functional as F
import json
from sklearn.metrics import f1_score
from collections import defaultdict
import copy
from transformers import T5Tokenizer, T5ForConditionalGeneration
from unidecode import unidecode
from sacremoses import MosesTokenizer, MosesDetokenizer
tokenizer = T5Tokenizer.from_pretrained('t5-large')
import logging
logging.basicConfig(level=logging.INFO)
logging.info('Start Logging')
def jload(fname):
t = open(fname,'r').read().replace('\u2013', '-').replace('~', '-').replace('`', "'").replace('\u2019', "'").replace('^', '')
return unidecode(t).split('\n')[:-1]
def prep_data(config, fname):
#prep data always has two steps, build the vocabulary first and then generate data samples
nsplit = int(config['gpus'])
train_raw = jload(config['train_file'])
train = [('Graph to Text: ' + x.split('||')[0].replace('_', ' '), x.split('||')[1]+' </s>') for x in train_raw]
dev_raw = jload(config['dev_file'])
dev = [('Graph to Text: ' + x.split('||')[0].replace('_', ' '), x.split('||')[1]+' </s>') for x in dev_raw]
test_raw = jload(config['test_file'])
test = [('Graph to Text: ' + x.replace('_', ' '), None) for x in test_raw]
print(len(train), len(dev), len(test))
sp = len(train)//nsplit
for i in range(nsplit):
torch.save({'train':train[i*sp:(i+1)*sp], 'dev':dev, 'test':test}, fname+str(i))
def pred_one(batch, model):
model.eval()
inp = tokenizer([x[0] for x in batch], return_tensors='pt', padding=True)['input_ids'].to(model.device)
pred = model.generate(input_ids=inp, max_len=50, num_beams=4)
return tokenizer.batch_decode(pred)
def train_g2t_one_step(batch, model, optimizer, config):
model.train()
optimizer.zero_grad()
inp = tokenizer([x[0] for x in batch], return_tensors='pt', padding=True)['input_ids'].to(model.device)
tar = tokenizer([x[1] for x in batch], return_tensors='pt', padding=True)['input_ids'].to(model.device)
loss = model(input_ids=inp, labels=tar)[0]
loss.backward()
for param in model.parameters():
if param.grad is not None:
torch.distributed.all_reduce(param.grad.data, op=torch.distributed.ReduceOp.SUM)
param.grad.data /= config['gpus']
nn.utils.clip_grad_norm_(model.parameters(), config['clip'])
optimizer.step()
return loss.item()
def eval_g2t(datas, model, demo_name='hyp.txt'):
model.eval()
hyp = []
with tqdm.tqdm(batch_it(datas, 2)) as tqb:
for i, batch in enumerate(tqb):
with torch.no_grad():
pred = pred_one(batch, model)
hyp.extend(pred)
mt = MosesTokenizer(lang='en')
md = MosesDetokenizer(lang='en')
wf_h = open(demo_name, 'w')
for i,h in enumerate(hyp):
wf_h.write(md.detokenize(mt.tokenize(str(h)))+'\n')
wf_h.close()
return 0.0
def batch_it(datas, batch_size):
r = []
ret = []
for x in datas:
r.append(x)
if len(r)==batch_size:
ret.append(r)
r = []
if len(r)>0:
ret.append(r)
return ret
def train(proc_id, devices, _type, config, fname='tmp_data.pt'):
random.seed(config['seed'])
torch.manual_seed(config['seed'])
np.random.seed(config['seed'])
torch.cuda.manual_seed_all(config['seed'])
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
dev_id = devices[proc_id]
device = torch.device(dev_id)
_d = torch.load(fname+str(proc_id))
train_d = _d['train']
dev_d = _d['dev']
test_d = _d['test']
port = 12346
dist_init_method = 'tcp://{master_ip}:{master_port}'.format(
master_ip='127.0.0.1', master_port=str(port))
torch.distributed.init_process_group(backend="nccl",
init_method=dist_init_method,
world_size=len(devices),
rank=dev_id)
model = T5ForConditionalGeneration.from_pretrained('t5-large')
model.config.update(model.config.task_specific_params['translation_en_to_de'])
model.to(device)
from transformers.optimization import get_cosine_schedule_with_warmup , get_linear_schedule_with_warmup
optimizer = torch.optim.Adam(model.parameters(), lr = config['lr'], weight_decay=config['weight_decay'])
# schedulerG2T = get_cosine_schedule_with_warmup(
# optimizer = optimizerG2T ,
# num_warmup_steps = 1500 ,
# num_training_steps = 3000 * config['main']['epoch'],
# )
losses = []
for i in range(0, config['epoch']):
with tqdm.tqdm(batch_it(train_d, config['batch_size'])) as tqb:
for j, batch in enumerate(tqb):
loss = train_g2t_one_step(batch, model, optimizer, config)
losses.append(loss)
tqb.set_postfix({'loss': np.mean(losses)})
logging.info('Epoch '+str(i))
if i%1==0 and proc_id==0:
torch.save(model.state_dict(), config['save']+'X'+str(i))
eval_g2t(test_d, model, 'hyp'+str(i)+'.txt')
torch.distributed.barrier()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--config', type=str, default='config.yaml')
args = parser.parse_args()
config = yaml.load(open(args.config, 'r'))
_config = copy.deepcopy(config)
random.seed(config['seed'])
torch.manual_seed(config['seed'])
np.random.seed(config['seed'])
torch.cuda.manual_seed_all(config['seed'])
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
fname = 'tmp_data.pt'
prep_data(config, fname)
devices = list(range(config['gpus']))
torch.multiprocessing.spawn(train, args=(devices, 'train', config, fname), nprocs=len(devices))
if __name__=='__main__':
main()