-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
155 lines (117 loc) · 6.86 KB
/
train.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
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
from config import LABEL_MAP
from model import Criterion, ISRScheduler
from utils import convert_ids_to_tokens, safe_loss, process_outputs, get_bow_labels
def train_generator(config, vocab, batch, model_G, model_D, optim_G, optim_D, seq_criterion, cls_criterion):
inputs, labels = batch
# rec loss
rec_logits, (_, _) = model_G(inputs, labels, targets=inputs[:, 1:])
rec_loss = seq_criterion(rec_logits, inputs[:, 1:], truncate=True)
# cyc loss
rev_logits, (rec_q, rev_samples) = model_G(inputs, 1-labels, targets=None)
rev_inputs = process_outputs(torch.argmax(rev_logits, dim=-1), vocab.eos, vocab.pad, vocab.sos)
cyc_logits, (rev_q, raw_samples) = model_G(rev_inputs, labels, targets=inputs[:, 1:])
cyc_loss = seq_criterion(cyc_logits, inputs[:, 1:], truncate=True)
# retrieval loss
cosloss = nn.CosineEmbeddingLoss(reduction="none")
targets = torch.ones(rev_q.size(0)).to(rev_q.device)
q_loss = cosloss(rec_q, rev_q, targets)
q_loss = safe_loss(q_loss)
q_loss = q_loss.mean()
# bow loss
rev_samples_bow = rev_samples.view(inputs.size(0), config.K, -1)
raw_samples_bow = raw_samples.view(inputs.size(0), config.K, -1)
rev_bow_labels = get_bow_labels(inputs, rev_samples_bow, vocab).unsqueeze(1).repeat(1, rev_logits.size(1), 1).view(-1, len(vocab))
cyc_bow_labels = get_bow_labels(rev_inputs, raw_samples_bow, vocab).unsqueeze(1).repeat(1, cyc_logits.size(1), 1).view(-1, len(vocab))
rev_bow_labels_num = rev_bow_labels.sum(dim=-1)
cyc_bow_labels_num = cyc_bow_labels.sum(dim=-1)
rev_bow_loss = (- F.log_softmax(rev_logits.view(-1, rev_logits.size(-1)), dim=-1) * rev_bow_labels).sum(dim=-1) / (rev_bow_labels_num+1e-2)
rev_bow_loss = safe_loss(rev_bow_loss.view(inputs.size(0), -1))
cyc_bow_loss = (- F.log_softmax(cyc_logits.view(-1, cyc_logits.size(-1)), dim=-1) * cyc_bow_labels).sum(dim=-1) / (cyc_bow_labels_num+1e-2)
cyc_bow_loss = safe_loss(cyc_bow_loss.view(inputs.size(0), -1))
bow_loss = (rev_bow_loss.mean() + cyc_bow_loss.mean()) / 2
# adv loss
rev_input_embeds = model_G.get_word_embedding(rev_logits)
class_logits = model_D(rev_input_embeds)
adv_loss = cls_criterion(class_logits, 1-labels)
losses = rec_loss + cyc_loss + adv_loss + q_loss + bow_loss
losses = losses.mean()
optim_G.zero_grad()
losses.backward()
torch.nn.utils.clip_grad_norm_(model_G.parameters(), config.max_grad_norm)
optim_G.step()
return losses.item(), rec_loss.item(), cyc_loss.item(), adv_loss.item(), q_loss.item(), bow_loss.item()
def train_disriminator(config, vocab, batch, model_G, model_D, optim_G, optim_D, seq_criterion, cls_criterion):
inputs, labels = batch
# rec
with torch.no_grad():
rec_logits, (rec_q, rec_samples) = model_G(inputs, labels, targets=inputs[:, 1:])
rev_logits, (rev_q, rev_samples) = model_G(inputs, 1-labels, targets=None)
rev_input_embeds = model_G.get_word_embedding(rev_logits)
raw_input_embeds = model_G.get_word_embedding(rec_logits)
golden_input_embedds = model_G.get_word_embedding(inputs[:, 1:])
golden_rec_embedds = model_G.get_word_embedding(rec_samples[:, 1:])
golden_rev_embedds = model_G.get_word_embedding(rev_samples[:, 1:])
rev_loss = cls_criterion(model_D(rev_input_embeds), torch.zeros_like(labels).fill_(2))
rec_loss = cls_criterion(model_D(raw_input_embeds), labels)
raw_loss = cls_criterion(model_D(golden_input_embedds), labels)
sample_raw_loss = cls_criterion(model_D(golden_rec_embedds), labels.repeat(config.K))
sample_rev_loss = cls_criterion(model_D(golden_rev_embedds), 1-labels.repeat(config.K))
dis_loss = raw_loss + rec_loss + rev_loss + sample_raw_loss + sample_rev_loss
optim_D.zero_grad()
dis_loss.backward()
torch.nn.utils.clip_grad_norm_(model_D.parameters(), config.max_grad_norm)
optim_D.step()
return dis_loss.item() / 5
def train(config, log, vocab, model_G, model_D, train_dataloader, dev_dataloader):
optim_G = torch.optim.Adam(model_G.parameters(), lr=0.0, betas=(0.9, 0.999))
optim_D = torch.optim.Adam(model_D.parameters(), lr=0.0, betas=(0.5, 0.999))
optim_G = ISRScheduler(optimizer=optim_G, warmup_steps=config.warmup_steps,
max_lr=config.max_lr, min_lr=config.min_lr, init_lr=config.init_lr, beta=0.75)
optim_D = ISRScheduler(optimizer=optim_D, warmup_steps=config.warmup_steps,
max_lr=config.max_lr, min_lr=config.min_lr, init_lr=config.init_lr, beta=0.5)
seq_criterion = Criterion(vocab.pad)
cls_criterion = Criterion(-1)
global_steps = 0
for epoch in range(config.epochs):
for step, batch in enumerate(train_dataloader):
if (global_steps) % config.update_step == 0:
log.info("Updating query embeddings.. ")
model_G.eval()
model_G.update_retrieval()
model_D.train()
model_G.train()
# train generator
losses, rec_loss, cyc_loss, adv_loss, cos_loss, bow_loss = \
train_generator(config, vocab, batch, model_G, model_D, optim_G, optim_D, seq_criterion, cls_criterion)
# train disriminator
for _ in range(config.d_step):
dis_loss = train_disriminator(config, vocab, batch, model_G, model_D, optim_G, optim_D, seq_criterion, cls_criterion)
if global_steps%20==0:
log.info(
"epoch {}, step {}/{}, losses: {:.2f}, rec_loss: {:.2f}, cyc_loss: {:.2f}, adv_loss: {:.2f}, cos_loss: {:.2f}, bow_loss: {:.2f}, dis_loss: {:.2f}, lr_G: {:.5f}, lr_D:{:.5f}".format(
epoch+1, step+1, len(train_dataloader), losses, rec_loss, cyc_loss, adv_loss, cos_loss, bow_loss, dis_loss, optim_G.rate(), optim_D.rate()
))
global_steps += 1
log.info("saving epoch {}...".format(epoch+1))
torch.save(model_G, os.path.join(config.ckpts_dir, "G_{}.pt".format(epoch+1)))
torch.save(model_D, os.path.join(config.ckpts_dir, "D_{}.pt".format(epoch+1)))
results = []
model_G.eval()
with torch.no_grad():
for loader in dev_dataloader.loaders:
outputs = []
for batch_test in loader:
outs, (_, _) = model_G(batch_test[0], 1 - batch_test[1])
outs = outs.argmax(dim=-1)
outs = convert_ids_to_tokens(outs, vocab)
outputs += outs
results.append(outputs)
for i, res in enumerate(results):
with open(os.path.join(config.out_dir, "epoch_{}.{}".format(epoch+1, LABEL_MAP[config.dataset][i])), "w", encoding="utf8") as f:
for line in res:
f.write(line + "\n")
return