-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathlevenshtein_en_de.py
165 lines (131 loc) · 7.55 KB
/
levenshtein_en_de.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
import torch
import torch.nn as nn
from torchtext import data, datasets
from transformer.optimizer import NoamOpt
from levenhtein_transformer.train import run_epoch
from levenhtein_transformer.criterion import LabelSmoothingLoss
from levenhtein_transformer.model import LevenshteinTransformerModel
from levenhtein_transformer.data import rebatch_and_noise, batch_size_fn, MyIterator
from levenhtein_transformer.validator import validate
from utils import save_model
from levenhtein_transformer.config import config
import wandb
BOS_WORD = '<s>'
EOS_WORD = '</s>'
BLANK_WORD = '<blank>'
UNK = '<unk>'
wandb.init(project="levenshtein_transformer")
wandb.config.update(config)
def main():
devices = list(range(torch.cuda.device_count()))
print('Selected devices: ', devices)
def tokenize_bpe(text):
return text.split()
SRC = data.Field(tokenize=tokenize_bpe, pad_token=BLANK_WORD, unk_token=UNK)
TGT = data.Field(tokenize=tokenize_bpe, init_token=BOS_WORD, unk_token=UNK,
eos_token=EOS_WORD, pad_token=BLANK_WORD)
train, val, test = datasets.WMT14.splits(exts=('.en', '.de'),
train='train.tok.clean.bpe.32000',
# train='newstest2014.tok.bpe.32000',
validation='newstest2013.tok.bpe.32000',
test='newstest2014.tok.bpe.32000',
fields=(SRC, TGT),
filter_pred=lambda x: len(vars(x)['src']) <= config['max_len'] and
len(vars(x)['trg']) <= config['max_len'],
root='./.data/')
print('Train set length: ', len(train))
wandb.config.update({'Train set length': len(train)})
# building shared vocabulary
TGT.build_vocab(train.src, train.trg, min_freq=config['min_freq'])
SRC.vocab = TGT.vocab
print('Source vocab length: ', len(SRC.vocab.itos))
print('Target vocab length: ', len(TGT.vocab.itos))
wandb.config.update({'Source vocab length': len(SRC.vocab), 'Target vocab length': len(TGT.vocab)})
pad_idx = TGT.vocab.stoi[BLANK_WORD]
bos_idx = TGT.vocab.stoi[BOS_WORD]
eos_idx = TGT.vocab.stoi[EOS_WORD]
unk_idx = TGT.vocab.stoi[UNK]
print(f'Indexes -- PAD: {pad_idx}, EOS: {eos_idx}, BOS: {bos_idx}, UNK: {unk_idx}')
train_iter = MyIterator(train, batch_size=config['batch_size'], device=torch.device(0), repeat=False,
sort_key=lambda x: (len(x.src), len(x.trg)), batch_size_fn=batch_size_fn, train=True)
valid_iter = MyIterator(val, batch_size=config['val_batch_size'], device=torch.device(0), repeat=False,
sort_key=lambda x: (len(x.src), len(x.trg)), batch_size_fn=batch_size_fn, train=False)
test_iter = MyIterator(test, batch_size=config['val_batch_size'], device=torch.device(0), repeat=False,
sort_key=lambda x: (len(x.src), len(x.trg)), batch_size_fn=batch_size_fn, train=False)
criterion = LabelSmoothingLoss(batch_multiplier=config['batch_multiplier'])
criterion.cuda()
# model = LevenshteinTransformerModel(len(SRC.vocab), len(TGT.vocab), n=1, PAD=pad_idx,
# BOS=bos_idx, EOS=eos_idx, UNK=unk_idx,
# criterion=criterion,
# d_model=256, d_ff=256, h=1,
# dropout=config['dropout'],
# input_dropout=config['input_dropout'])
model = LevenshteinTransformerModel(len(SRC.vocab), len(TGT.vocab),
n=config['num_layers'],
h=config['attn_heads'],
d_model=config['model_dim'],
dropout=config['dropout'],
input_dropout=config['input_dropout'],
d_ff=config['ff_dim'],
criterion=criterion,
PAD=pad_idx, BOS=bos_idx, EOS=eos_idx, UNK=unk_idx)
# weight tying
model.src_embed[0].lookup_table.weight = model.tgt_embed[0].lookup_table.weight
model.generator.lookup_table.weight = model.tgt_embed[0].lookup_table.weight
model.decoder.out_layer.lookup_table.weight = model.tgt_embed[0].lookup_table.weight
model.cuda()
model_size = model.src_embed[0].d_model
print('Model created with size of', model_size)
wandb.config.update({'Model size': model_size})
# make the inner model functions available from the DataParallel wrapper
class MyDataParallel(nn.DataParallel):
def __getattr__(self, name):
try:
return super(MyDataParallel, self).__getattr__(name)
except AttributeError:
return getattr(self.module, name)
model_par = MyDataParallel(model, device_ids=devices)
model_opt = NoamOpt(warmup_init_lr=config['warmup_init_lr'], warmup_end_lr=config['warmup_end_lr'],
warmup_updates=config['warmup'],
min_lr=config['min_lr'],
optimizer=torch.optim.Adam(model.parameters(),
lr=0,
weight_decay=config['weight_decay'],
betas=(config['beta_1'], config['beta_2']),
eps=config['epsilon']))
wandb.watch(model)
current_steps = 0
epoch = 0
while True:
# training model
print('Epoch ', epoch)
wandb.log({'Epoch': epoch}, commit=False)
model_par.train()
loss, steps = run_epoch((rebatch_and_noise(b, pad=pad_idx, bos=bos_idx, eos=eos_idx) for b in train_iter),
model=model_par,
opt=model_opt,
steps_so_far=current_steps,
batch_multiplier=config['batch_multiplier'],
logging=True,
train=True)
current_steps += steps
if epoch >= 2:
save_model(model=model, optimizer=model_opt.optimizer, loss=loss, src_field=SRC, tgt_field=TGT,
updates=current_steps, epoch=epoch, prefix=f'lev_t_epoch_{epoch}___')
# calculating validation bleu score
model_par.eval()
bleu = validate(model=model_par,
iterator=(rebatch_and_noise(b, pad=pad_idx, bos=bos_idx, eos=eos_idx) for b in valid_iter),
SRC=SRC, TGT=TGT, EOS_WORD=EOS_WORD, bos=bos_idx, eos=eos_idx, pad=pad_idx,
max_decode_iter=min(epoch + 1, config['max_decode_iter']), logging=False)
wandb.log({'Epoch bleu score': bleu}, commit=False)
if current_steps > config['max_step']:
break
epoch += 1
test_bleu = validate(model=model_par,
iterator=(rebatch_and_noise(b, pad=pad_idx, bos=bos_idx, eos=eos_idx) for b in test_iter),
SRC=SRC, TGT=TGT, EOS_WORD=EOS_WORD, bos=bos_idx, eos=eos_idx, pad=pad_idx,
max_decode_iter=config['max_decode_iter'], logging=False, is_test=True)
print(f"Test Bleu score: {test_bleu}")
wandb.config.update({'Test bleu score': test_bleu})
main()