-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_dual_mode.py
141 lines (123 loc) · 6.27 KB
/
train_dual_mode.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
import os
import math
import random
import argparse
import h5py
import torch
import torch.nn as nn
import numpy as np
from torch.utils.data import DataLoader
from tqdm import tqdm
from model import *
from data_load import *
import scoring
import subprocess
from Loss import Loss_kd
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = False
def get_lr(optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
def get_output(outputs, seq_len):
output_ = 0
for i in range(len(seq_len)):
length = seq_len[i]
output = outputs[i, :length, :]
if i == 0:
output_ = output
else:
output_ = torch.cat((output_, output), dim=0)
return output_
def main():
parser = argparse.ArgumentParser(description='paras for making data')
parser.add_argument('--dim', type=int, help='dim of input features', default=80)
parser.add_argument('--model', type=str, help='model name for saving', default='XSA_E2E')
parser.add_argument('--train', type=str, help='training data, in .txt')
parser.add_argument('--batch', type=int, help='batch size', default=32)
parser.add_argument('--optim', type=str, help='optimizer', default='warmcosine')
parser.add_argument('--warmup', type=int, help='num of steps', default=24000)
parser.add_argument('--epochs', type=int, help='num of epochs', default=20)
parser.add_argument('--lang', type=int, help='num of language classes', default=14)
parser.add_argument('--lr', type=float, help='initial learning rate', default=0.0001)
parser.add_argument('--device', type=int, help='Device name', default=0)
parser.add_argument('--seed', type=int, help='seed', default=0)
parser.add_argument('--temperature', type=int, help='temperature', default=2)
parser.add_argument('--window', type=str, help='fix or random window in student model', default='fix')
parser.add_argument('--winlen', type=int, help='window length', default=15)
parser.add_argument('--alpha', type=float, help='importance of teacher (full mode)', default=0.33)
parser.add_argument('--beta', type=float, help='importance of student (short mode)', default=0.33)
args = parser.parse_args()
setup_seed(args.seed)
device = torch.device('cuda:{}'.format(args.device) if torch.cuda.is_available() else 'cpu')
model = X_Transformer_E2E_LID(input_dim=args.dim,
feat_dim=64,
d_k=64,
d_v=64,
d_ff=2048,
n_heads=8,
dropout=0.1,
n_lang=args.lang,
max_seq_len=10000)
model.to(device)
train_txt = args.train
train_set = RawFeatures(train_txt)
train_data = DataLoader(dataset=train_set,
batch_size=args.batch,
pin_memory=False,
num_workers=16,
shuffle=True,
collate_fn=collate_fn_atten)
loss_func_CRE = nn.CrossEntropyLoss().to(device)
KD_loss_func = nn.KLDivLoss(reduction='batchmean').to(device)
total_step = len(train_data)
optimizer = args.optim
if optimizer == 'noam':
optimizer = torch.optim.Adam(model.parameters(), betas=(0.9, 0.98), eps=1e-9, lr=0.1)
warm_up_with_noam = lambda step: (512 ** (-0.5)) * min((step + 1) ** (-0.5), (step + 1) * args.warmup ** (-1.5))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=warm_up_with_noam)
elif optimizer == 'warmcosine':
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
warm_up_with_cosine_lr = lambda step: step / args.warmup \
if step <= args.warmup \
else 0.5 * (math.cos((step - args.warmup) / (args.epochs * total_step - args.warmup) * math.pi) + 1)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=warm_up_with_cosine_lr)
elif optimizer == 'cosine':
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs * total_step)
# ====Train=====
for epoch in tqdm(range(args.epochs)):
model.train()
for step, (utt, labels, seq_len) in enumerate(train_data):
utt_ = utt.to(device=device, dtype=torch.float)
atten_mask = get_atten_mask(seq_len, utt_.size(0))
atten_mask = atten_mask.to(device=device)
atten_mask_student = get_atten_mask_student(seq_len, utt_.size(0),
mask_type=args.window, win_len=args.winlen)
atten_mask_student = atten_mask_student.to(device=device)
labels = labels.to(device=device, dtype=torch.long)
# Forward pass
outputs = model(utt_, seq_len, atten_mask)
outputs_student = model(utt_, seq_len, atten_mask_student)
loss_teacher = loss_func_CRE(outputs, labels)
loss_student = loss_func_CRE(outputs_student, labels)
loss_kd = KD_loss_func(F.log_softmax(outputs_student / args.temperature, dim=1),
F.softmax(outputs / args.temperature, dim=1)) * (args.temperature * args.temperature)
loss = args.alpha * loss_teacher + args.beta * loss_student + (1 - args.alpha - args.beta) * loss_kd
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 1000 == 0:
print("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f} loss_T: {:.4f} loss_S: {:.4f}, loss_KD: {:.4f}".
format(epoch + 1, args.epochs, step + 1, total_step, loss.item(), loss_teacher.item(),
loss_student.item())), loss_kd.item()))
scheduler.step()
print(get_lr(optimizer))
if epoch >= args.epochs - 5:
torch.save(model.state_dict(), '/home/hexin/Desktop/models/' + '{}_epoch_{}.ckpt'.format(args.model, epoch))
if __name__ == "__main__":
main()