FilesExpand file tree

master

/

train.py

Copy path

Blame

More file actions

Blame

More file actions

Latest commit

 

History

History

History

195 lines (153 loc) · 5.66 KB

master

/

train.py

Copy path

Top

File metadata and controls

• Code
• Blame

195 lines (153 loc) · 5.66 KB

Raw

Copy raw file

Download raw file

Open symbols panel

Edit and raw actions

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

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

# -*- coding: utf-8 -*-

import argparse

import math

import os

import dill

from collections import OrderedDict

from tqdm import tqdm

from torchtext import data

from torchtext import datasets

from torchtext.vocab import Vectors

import torch

import torch.nn as nn

import torch.optim as optim

from options import train_opts

from options import model_opts

from model import Seq2seqAttn

class Trainer(object):

def __init__(

self, model, criterion, optimizer, scheduler, clip):

self.model = model

self.criterion = criterion

self.optimizer = optimizer

self.scheduler = scheduler

self.clip = clip

self.n_updates = 0

def get_lr(self):

return self.optimizer.param_groups[0]['lr']

def step(self, samples, tf_ratio):

self.optimizer.zero_grad()

bsz = samples.src.size(1)

outs = self.model(samples.src, samples.tgt, tf_ratio)

loss = self.criterion(outs.view(-1, outs.size(2)), samples.tgt.view(-1))

if self.model.training:

loss.backward()

nn.utils.clip_grad_norm_(self.model.parameters(), self.clip)

self.optimizer.step()

self.n_updates += 1

return loss

def save_model(save_vars, filename):

model_path = os.path.join(args.savedir, filename)

torch.save(save_vars, model_path)

def save_vocab(savedir, fields):

name, field = fields

save_path = os.path.join(savedir, f"{name}_vocab.txt")

with open(save_path, 'w') as fout:

for w in field.vocab.itos:

fout.write(w + '\n')

def save_field(savedir, fields):

name, field = fields

save_path = os.path.join(savedir, f"{name}.field")

with open(save_path, 'wb') as fout:

dill.dump(field, fout)

def main(args):

device = torch.device('cuda' if args.gpu else 'cpu')

# load data and construct vocabulary dictionary

SRC = data.Field(lower=True)

TGT = data.Field(lower=True, eos_token='<eos>')

fields = [('src', SRC), ('tgt', TGT)]

train_data = data.TabularDataset(

path=args.train,

format='tsv',

fields=fields,

)

valid_data = data.TabularDataset(

path=args.valid,

format='tsv',

fields=fields,

)

SRC.build_vocab(train_data, min_freq=args.src_min_freq)

TGT.build_vocab(train_data, min_freq=args.tgt_min_freq)

if not os.path.exists(args.savedir):

os.mkdir(args.savedir)

# save field and vocabulary

for field in fields:

save_field(args.savedir, field)

save_vocab(args.savedir, field)

# set iterator

train_iter, valid_iter = data.BucketIterator.splits(

(train_data, valid_data),

batch_size=args.batch_size,

sort_within_batch=True,

sort_key= lambda x: len(x.src),

repeat=False,

device=device

)

model = Seq2seqAttn(args, fields, device).to(device)

print(model)

print('')

criterion = nn.CrossEntropyLoss(ignore_index=TGT.vocab.stoi['<pad>'])

optimizer = optim.SGD(model.parameters(), lr=args.lr)

scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min')

trainer = Trainer(model, criterion, optimizer, scheduler, args.clip)

epoch = 1

max_epoch = args.max_epoch or math.inf

max_update = args.max_update or math.inf

best_loss = math.inf

while epoch < max_epoch and trainer.n_updates < max_update \

and args.min_lr < trainer.get_lr():

# training

with tqdm(train_iter, dynamic_ncols=True) as pbar:

train_loss = 0.0

trainer.model.train()

for samples in pbar:

bsz = samples.src.size(1)

loss = trainer.step(samples, args.tf_ratio)

train_loss += loss.item()

# setting of progressbar

pbar.set_description(f"epoch {str(epoch).zfill(3)}")

progress_state = OrderedDict(

loss=loss.item(),

ppl=math.exp(loss.item()),

bsz=len(samples),

lr=trainer.get_lr(),

clip=args.clip,

num_updates=trainer.n_updates)

pbar.set_postfix(progress_state)

train_loss /= len(train_iter)

print(f"| epoch {str(epoch).zfill(3)} | train ", end="")

print(f"| loss {train_loss:.{4}} ", end="")

print(f"| ppl {math.exp(train_loss):.{4}} ", end="")

print(f"| lr {trainer.get_lr():.1e} ", end="")

print(f"| clip {args.clip} ", end="")

print(f"| num_updates {trainer.n_updates} |")

# validation

valid_loss = 0.0

trainer.model.eval()

for samples in valid_iter:

bsz = samples.src.size(1)

loss = trainer.step(samples, tf_ratio=0.0)

valid_loss += loss.item()

valid_loss /= len(valid_iter)

print(f"| epoch {str(epoch).zfill(3)} | valid ", end="")

print(f"| loss {valid_loss:.{4}} ", end="")

print(f"| ppl {math.exp(valid_loss):.{4}} ", end="")

print(f"| lr {trainer.get_lr():.1e} ", end="")

print(f"| clip {args.clip} ", end="")

print(f"| num_updates {trainer.n_updates} |")

# saving model

save_vars = {"train_args": args,

"state_dict": model.state_dict()}

if valid_loss < best_loss:

best_loss = valid_loss

save_model(save_vars, 'checkpoint_best.pt')

save_model(save_vars, "checkpoint_last.pt")

# update

trainer.scheduler.step(valid_loss)

epoch += 1

if __name__ == '__main__':

parser = argparse.ArgumentParser('')

train_opts(parser)

model_opts(parser)

args = parser.parse_args()

main(args)