-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_utils.py
161 lines (128 loc) · 5.3 KB
/
train_utils.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
160
'''
@Description:
@Author: bairongz ([email protected])
@Date: 2020-01-03 23:05:17
@LastEditTime : 2020-01-05 14:24:40
'''
import torch
from torch import nn
from model import *
from tqdm import tqdm
import pickle
def sanity_check(args):
pass
def show_training_profile(args, model):
print(args)
print(model)
def save_ckpt(path, args, model, optim, lang):
"""save checkpoint for warmstart, evaluation.
"""
print("saving model to %s" % path)
torch.save({
"args": args,
'model_state_dict': model.state_dict(),
# 'optim_state_dict': optim.state_dict(),
"lang": lang
}, path)
def load_ckpt(path, with_optim):
"""load checkpoint for warm start, evaluation.
"""
print("loading model from %s" % path)
ckpt = torch.load(path)
args = ckpt["args"]
print(args)
model = load_model(args)
model.load_state_dict(ckpt["model_state_dict"])
optim=None
if with_optim:
optim.load_state_dict(ckpt["optim_state_dict"])
lang = ckpt["lang"]
return args, model, optim, lang
def load_model(args):
if args.model_type == "trs":
print("creating vanilla Transformer model...")
model = Transformer(args.n_src_vocab,
args.d_model,
args.n_head,
args.n_enc_layers,
args.n_dec_layers,
args.d_ff,
share_word_embedding=args.share_embed,
n_dec_vocab=args.n_trg_vocab,
dropout=args.dropout)
elif args.model_type == "ut":
print("creating Universal Transformer model...")
if args.act:
print("Use ACT type: %s, epslion: %f" % (args.act, args.act_eps))
model = UniversalTransformer(args.n_src_vocab,
args.d_model,
args.n_head,
args.n_enc_layers,
args.n_dec_layers,
args.d_ff,
share_word_embedding=args.share_embed,
n_dec_vocab=args.n_trg_vocab,
act=args.act,
epslion=args.act_eps,
)
else:
raise ValueError("unknown model_type: %s" % args.model_type)
return model
def init_weight(args, model):
def weight_init(m):
init_method = None
if args.init == "normal":
init_method = torch.nn.init.normal_
elif args.init=="xavier_uniform":
init_method = torch.nn.init.xavier_uniform_
elif args.init == "xavier_normal":
init_method = torch.nn.init.xavier_normal_
elif args.init == "kaiming_uniform":
init_method = torch.nn.init.kaiming_uniform_
elif args.init == "kaiming_normal":
init_method = torch.nn.init.kaiming_normal_
elif args.init == "orthogonal":
init_method = torch.nn.init.orthogonal_
else:
raise ValueError("Unknown weight initialization method: %s"%args.init)
if isinstance(m, (nn.Linear)):
init_method(m.weight.data)
if m.bias is not None and args.init not in ["xavier_uniform", "xavier_normal", "kaiming_normal", "kaiming_uniform", "orthogonal"]:
init_method(m.bias.data)
model.apply(weight_init)
def train_step(model, batch, loss_fn, optim, norm=0.5, act_loss_weight=0, device="cpu"):
model.train()
src, trg, src_pos, trg_pos, trg_mask, src_key_padding_mask, trg_key_padding_mask, memory_key_padding_mask \
= list(map(lambda x: x.to(device), batch))
dec_src = trg[:-1].detach().contiguous()
trg = trg[1:].detach().contiguous()
logits = model(src, dec_src, src_pos, trg_pos, trg_mask=trg_mask,
src_key_padding_mask=src_key_padding_mask,
trg_key_padding_mask=trg_key_padding_mask,
memory_key_padding_mask=memory_key_padding_mask
)
loss = loss_fn(logits, trg.view(-1))
if isinstance(model, UniversalTransformer) and act_loss_weight:
loss += act_loss_weight * model.act_loss()
optim.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), norm)
optim.step()
return loss.item()
def valid_step(model, valid_iter, loss_fn, device="cpu"):
model.eval()
loss_record = []
with torch.no_grad():
for batch in valid_iter:
src, trg, src_pos, trg_pos, trg_mask, src_key_padding_mask, trg_key_padding_mask, memory_key_padding_mask \
= list(map(lambda x: x.to(device), batch))
dec_src = trg[:-1].detach().contiguous()
trg = trg[1:].detach().contiguous()
logits = model(src, dec_src, src_pos, trg_pos, trg_mask=trg_mask,
src_key_padding_mask=src_key_padding_mask,
trg_key_padding_mask=trg_key_padding_mask,
memory_key_padding_mask=memory_key_padding_mask
)
loss = loss_fn(logits, trg.view(-1))
loss_record.append(loss.item())
return np.mean(loss_record)