open_stt_e2e/train_rnnt.py at master · 1ytic/open_stt_e2e · GitHub

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
import sys
import torch
import numpy as np
import torch.nn as nn
from torch.optim.lr_scheduler import StepLR
from torch.utils.tensorboard import SummaryWriter
from torch_edit_distance import remove_blank, AverageWER, AverageCER
from data import Labels, split_train_dev_test
from model import Transducer
from utils import AverageMeter
from warp_rnnt import rnnt_loss

torch.backends.cudnn.benchmark = True
torch.manual_seed(1)
np.random.seed(1)

labels = Labels()

blank = torch.tensor([labels.blank()], dtype=torch.int).cuda()
space = torch.tensor([labels.space()], dtype=torch.int).cuda()

model = Transducer(128, len(labels), 512, 256, am_layers=3, lm_layers=3, dropout=0.3,
                   am_checkpoint='runs/Dec14_21-53-45_ctc_bs32x4_gn200/model20.bin',
                   lm_checkpoint='runs/Dec09_22-04-47_lm_bptt8_bs64_gn1_do0.3/model10.bin')
model.cuda()

train, dev, test = split_train_dev_test(
    '/open-stt-e2e/data/',
    labels, model.am.conv, batch_size=32
)

parameters = [
    {'params': model.fc.parameters(), 'lr': 3e-5},
    {'params': model.am.parameters(), 'lr': 3e-5},
    {'params': model.lm.parameters(), 'lr': 3e-5}
]

optimizer = torch.optim.Adam(parameters, weight_decay=1e-5)
scheduler = StepLR(optimizer, step_size=2000, gamma=0.99)

beta = 0.5

step = 0
writer = SummaryWriter(comment='_rnnt_bs32x4_gn200_beta0.5')

for epoch in range(1, 11):

    train.shuffle(epoch)

    model.train()

    err = AverageMeter('Loss/train')
    ent = AverageMeter('Entropy/train')
    grd = AverageMeter('Gradient/train')

    optimizer.zero_grad()

    for xs, ys, xn, yn in train:

        step += 1

        zs, xs, xn = model(xs, ys.t(), xn, yn)

        loss1 = rnnt_loss(zs, ys, xn, yn, average_frames=False, reduction="mean")

        loss2 = -(zs.exp() * zs).sum(dim=-1).mean()

        loss = loss1 - beta * loss2

        loss.backward()

        err.update(loss1.item())
        ent.update(loss2.item())

        writer.add_scalar(err.title + '/steps', loss1.item(), step)
        writer.add_scalar(ent.title + '/steps', loss2.item(), step)

        if step % 4 > 0:
            continue

        grad_norm = nn.utils.clip_grad_norm_(model.parameters(), 200)

        optimizer.step()
        scheduler.step()

        optimizer.zero_grad()

        grd.update(grad_norm)

        writer.add_scalar(grd.title + '/steps', grad_norm, step)

        train.set_description('Epoch %d %s %s %s' % (epoch, err, ent, grd))

    model.eval()

    for i, lr in enumerate(scheduler.get_lr()):
        writer.add_scalar('LR/%d' % i, lr, epoch)

    err.summary(writer, epoch)
    ent.summary(writer, epoch)
    grd.summary(writer, epoch)

    err = AverageMeter('Loss/test')
    ent = AverageMeter('Entropy/test')
    cer = AverageCER(blank, space)
    wer = AverageWER(blank, space)

    with torch.no_grad():

        temperature = 3
        prediction = []
        prior = 0

        for xs, ys, xn, yn in dev:

            xs, xn = model.forward_acoustic(xs, xn)

            xs = model.greedy_decode(xs, argmax=False)

            xs = xs.exp().view(-1, len(labels))

            prediction.append(xs.argmax(1).cpu())
            prior += xs.sum(dim=0)

            dev.set_description('Epoch %d Prior %.5f' % (epoch, prior.std().item()))

        prediction = torch.cat(prediction)
        prior = (prior / prediction.size(0)).log() / temperature

        writer.add_histogram('Prediction', prediction[prediction != labels.blank()], epoch)
        writer.add_histogram('Prior', prior, epoch)

        for xs, ys, xn, yn in test:

            zs, xs, xn = model(xs, ys.t(), xn, yn)

            loss1 = rnnt_loss(zs, ys, xn, yn, average_frames=False, reduction="mean")

            loss2 = -(zs.exp() * zs).sum(dim=-1).mean()

            xs = model.greedy_decode(xs, prior)

            err.update(loss1.item())
            ent.update(loss2.item())

            remove_blank(xs, xn, blank)

            cer.update(xs, ys, xn, yn)
            wer.update(xs, ys, xn, yn)

            test.set_description('Epoch %d %s %s %s %s' % (epoch, err, ent, cer, wer))

    sys.stderr.write('\n')

    err.summary(writer, epoch)
    ent.summary(writer, epoch)
    cer.summary(writer, epoch)
    wer.summary(writer, epoch)

    writer.flush()

    torch.save(model.state_dict(), writer.log_dir + '/model%d.bin' % epoch)

writer.close()