-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathreplace_sign_seq2seq.py
159 lines (137 loc) · 7.11 KB
/
replace_sign_seq2seq.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
import argparse
import os
import pickle
import pandas as pd
import torch
from torchtext.data.utils import get_tokenizer
from data.util import get_seq2seq_dataloader
from models.layers.util import random_flip_sign
from models.losses.sign_loss import SignLoss
from models.util import seed_everything
from trainer.nmt_seq2seq import EncDecEvaluator
def replace_sign(encoder, decoder, enc_optimizer, dec_optimizer, device, epochs):
converged = False
for _ in range(epochs):
encoder.train()
decoder.train()
enc_optimizer.zero_grad()
dec_optimizer.zero_grad()
# reset sign loss
for m in encoder.modules():
if isinstance(m, SignLoss):
m.reset()
for m in decoder.modules():
if isinstance(m, SignLoss):
m.reset()
kh = encoder.get_signature(reduce=False)
for m in encoder.modules():
if isinstance(m, SignLoss):
m.add(kh)
kh = decoder.get_signature(reduce=False)
for m in decoder.modules():
if isinstance(m, SignLoss):
m.add(kh)
sign_loss = torch.tensor(0.).to(device)
# add up sign loss
for m in encoder.modules():
if isinstance(m, SignLoss):
sign_loss += m.loss
for m in decoder.modules():
if isinstance(m, SignLoss):
sign_loss += m.loss
# print(f'Sign Loss: {sign_loss.item()}')
if sign_loss.item() < 0.0001:
converged = True
sign_loss.backward()
enc_optimizer.step()
dec_optimizer.step()
if converged:
break
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--test-file', type=str, dest='test_file',
default='/datadrive/rnn-ipr/wmt14-enfr/test/newstest2014-tokenized',
help='test file')
parser.add_argument('--output-dir', type=str, dest='output_dir', help='output folder')
parser.add_argument('--src', type=str, dest='src', default='en', help='source language')
parser.add_argument('--trg', type=str, dest='trg', default='fr', help='target language')
parser.add_argument('--src-vocab-path', type=str, dest='src_vocab_path', default='./outputs/enfr_en_vocab.pickle',
help='path to src vocab')
parser.add_argument('--trg-vocab-path', type=str, dest='trg_vocab_path', default='./outputs/enfr_fr_vocab.pickle',
help='path to trg vocab')
parser.add_argument('--seed', type=int, dest='seed', default=1234, help='seed for experiment')
parser.add_argument('--epochs', type=int, dest='epochs', default=500, help='number of epochs')
parser.add_argument('--batch-size', type=int, dest='batch_size', default=256, help='batch size per steps')
parser.add_argument('--max-sentence-length', type=int, dest='max_sentence_length', default=15,
help='max sentence length')
parser.add_argument('--reverse-input', action='store_true', dest='reverse_input', default=True,
help='reverse input sequence')
parser.add_argument('--pretrained-path', type=str, dest='pretrained_path', help='path to saved pretrained model',
required=True)
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
seed = args.seed
seed_everything(seed)
batch_size = args.batch_size
max_vocab = 15000
epochs = args.epochs
test_src_file = '{}.{}'.format(args.test_file, args.src)
test_trg_file = '{}.{}'.format(args.test_file, args.trg)
max_sentence_length = args.max_sentence_length
reverse_input = args.reverse_input
save_dir = args.output_dir
os.makedirs(save_dir, exist_ok=True)
with open(args.src_vocab_path, 'rb') as f:
src_vocab = pickle.load(f)
with open(args.trg_vocab_path, 'rb') as f:
trg_vocab = pickle.load(f)
num_words = min(max_vocab, len(src_vocab.itos))
num_words_outputs = min(max_vocab, len(trg_vocab.itos))
trg_pad_idx = trg_vocab.stoi['<pad>']
trg_eos_idx = trg_vocab.stoi['<eos>']
trg_sos_idx = trg_vocab.stoi['<sos>']
with open(os.path.join(args.pretrained_path, 'keyed_kwargs_{}.pickle'.format(seed)), 'rb') as f:
keyed_kwargs = pickle.load(f)
# loading dataset
test_dataloader, _, _ = get_seq2seq_dataloader(test_src_file, test_trg_file, in_vocab=src_vocab,
out_vocab=trg_vocab,
filters='•',
in_tokenizer=get_tokenizer(None, 'en'),
out_tokenizer=get_tokenizer(None, 'fr'),
batch_size=batch_size, max_vocab=max_vocab, shuffle=False,
max_sentence_length=max_sentence_length, test_size=None,
reverse_input=reverse_input
)
if os.path.isfile(os.path.join(args.pretrained_path, 'trigger_dataloader_{}.pth'.format(seed))):
print('found trigger dataset')
trigger_dataloader = torch.load(os.path.join(args.pretrained_path, 'trigger_dataloader_{}.pth'.format(seed)))
else:
trigger_dataloader = None
res = []
for perc in [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]:
encoder = torch.load(os.path.join(args.pretrained_path, 'encoder_{}.pth'.format(seed)))
decoder = torch.load(os.path.join(args.pretrained_path, 'decoder_{}.pth'.format(seed)))
# remove sign loss
encoder.gru.sign_loss = None
decoder.gru.sign_loss = None
enc_old_signature = torch.sign(encoder.get_signature().cpu().detach())
dec_old_signature = torch.sign(decoder.get_signature().cpu().detach())
enc_new_signature = random_flip_sign(enc_old_signature.clone(), perc).to(device)
dec_new_signature = random_flip_sign(dec_old_signature.clone(), perc).to(device)
# new sign loss with no regularizing
encoder.gru.sign_loss = SignLoss(1.0, enc_new_signature, regularize=False)
decoder.gru.sign_loss = SignLoss(1.0, dec_new_signature, regularize=False)
enc_optimizer = torch.optim.Adam(encoder.parameters())
dec_optimizer = torch.optim.Adam(decoder.parameters())
replace_sign(encoder, decoder, enc_optimizer, dec_optimizer, device, epochs)
print('*' * 50)
print(f'Evaluating with {perc} flipped sign:')
evaluator = EncDecEvaluator(encoder, decoder, device, trg_vocab)
te = evaluator.evaluate_bleu(test_dataloader, use_key=True)
if trigger_dataloader is not None:
tri = evaluator.evaluate_bleu(trigger_dataloader, use_key=True)
res.append({'flip_perc': perc, 'test_bleu': te['bleu_score'], 'trigger_bleu': tri['bleu_score']})
else:
res.append({'flip_perc': perc, 'test_bleu': te['bleu_score']})
train_df = pd.DataFrame(res)
train_df.to_csv(os.path.join(save_dir, 'results_{}.csv'.format(seed)))