-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
225 lines (192 loc) · 7.69 KB
/
main.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
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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
import torch
import time
from torch import Tensor
from rich.progress import Progress
from torch.utils.data import DataLoader, random_split
from dataset import PoetryData
from model import PoetryNet
batch_size = 64
lr = 0.0001
class PoetryGen:
def __init__(self) -> None:
# self.device = (
# torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")
# )
self.device = torch.device("cuda:0")
self.dataset = PoetryData(self.device, max_lines=50000, token_length=12)
self.vocab_size = self.dataset.vocab_size
train_data, test_data = random_split(
self.dataset, [len(self.dataset) - 1000, 1000]
)
self.train_dataloader = DataLoader(train_data, batch_size, True)
self.test_dataloader = DataLoader(test_data, batch_size, True)
self.net = PoetryNet(self.vocab_size, self.device, embed_size=512).to(
self.device
)
self.optimizer = torch.optim.Adam(self.net.parameters(), lr)
self.optimizer_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
self.optimizer, 256
)
self.loss_f = torch.nn.CrossEntropyLoss(ignore_index=2)
self.loaded_checkpoint_file = None
self.epoch = 0
self.progress = Progress()
import glob
files = glob.glob("checkpoint-*.pth")
for i, file in enumerate(files):
print(f"{i}> {file}")
if files:
t = input(
"choose check point to load, default is the last one, n to unload>"
)
if t == "":
t = -1
if t != "n":
self.load_checkpoint(files[int(t)])
def save_checkpoint(self):
file_name = (
self.loaded_checkpoint_file
or f'checkpoint-{time.strftime("%y%m%d-%H%M")}.pth'
)
with open(file_name, "wb") as file:
torch.save(
{
"net_state": self.net.state_dict(),
"optimizer_state": self.optimizer.state_dict(),
"epoch": self.epoch,
},
file,
)
print(f"save check point to {file_name}")
self.loaded_checkpoint_file = file_name
def load_checkpoint(self, file: str):
ckpt = torch.load(file)
self.net.load_state_dict(ckpt["net_state"])
self.optimizer.load_state_dict(ckpt["optimizer_state"])
self.epoch = ckpt["epoch"]
self.loaded_checkpoint_file = file
self.optimizer_scheduler.last_epoch = self.epoch
print(f"loaded check point: {file}, epoch: {self.epoch}")
def generate_one(self, pre_sentence: str, start_words: str = ""):
self.net.eval()
start_words_token = [0]
start_words_token.extend(self.dataset.word2idx[x] for x in start_words)
src = self.dataset.word2token(pre_sentence).unsqueeze(0)
tgt = torch.LongTensor([start_words_token]).to(self.device)
memo = self.net.encode(src)
res = []
for i in range(12):
out = self.net.decode(tgt, memo)
next_word = out.argmax(2)
if next_word[0][-1] == 1:
break
# tgt = torch.cat((tgt, ))
res.append(next_word[0][-1].item())
tgt = torch.cat((tgt, next_word[:, -1:]), 1)
return start_words + self.dataset.token2word(res)
def generate(self, num_sentence: int, pre_style: str):
res = []
for i in range(num_sentence):
s = self.generate_one(pre_style if not res else res[-1])
res.append(s)
return "/".join(res)
def generate_by_start(self, start_words: str, pre_style: str) -> str:
"""generate sentence by start words
Args:
start_words (str): start words for ever sentence, Divide by /, for example: 我/你/他/你
pre_style(str): style for the poem
Returns:
str: the result
"""
res = []
start_words_l = start_words.split("/")
if not start_words_l:
return ""
for i, s in enumerate(start_words_l):
t = self.generate_one(pre_style if not res else res[-1], s)
res.append(t)
return "/".join(res)
def forward_net(self, src: Tensor, tgt: Tensor):
src, tgt = src.to(self.device), tgt.to(self.device)
src_mask = (src == 2).to(self.device)
dec_tgt = tgt[:, :-1]
dec_tgt_mask = (dec_tgt == 2).to(self.device)
tgt_mask = torch.nn.Transformer.generate_square_subsequent_mask(
dec_tgt.size(1), self.device
)
out = self.net.forward(src, dec_tgt, tgt_mask, src_mask, dec_tgt_mask)
return out
def train_epoch(self):
self.net.train()
train_progress = self.progress.add_task(
description="Train Epoch", total=len(self.train_dataloader)
)
# ignore <pad> which index is 2
loss_f = self.loss_f # torch.nn.CrossEntropyLoss(ignore_index=2)
voacb_size = self.dataset.vocab_size
len_data = len(self.train_dataloader)
loss_all = 0
for i, (src, tgt) in enumerate(self.train_dataloader):
out = self.forward_net(src, tgt)
# out = torch.max(out, 2).indices
loss = loss_f.forward(out.reshape(-1, voacb_size), tgt[:, 1:].flatten())
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
self.progress.update(
train_progress,
advance=1,
description=f"{i}/{len_data} loss={loss.item():.4f}",
)
loss_all += loss.item()
# if i % 40 == 0:
# print(f"epoch={self.epoch} loss={loss.item():.4f} {i}")
# # print(self.dataset.token2word(tgt[0]))
# # print(self.dataset.token2word(out[0].argmax(1).tolist()))
self.optimizer_scheduler.step()
self.progress.remove_task(train_progress)
self.progress.print(
f"train epoch={self.epoch} average loss={loss_all/len_data:.4f} lr={self.optimizer_scheduler.get_lr()}"
)
def evaluation(self):
self.net.eval()
loss_f = self.loss_f
voacb_size = self.dataset.vocab_size
loss_a = 0
with torch.no_grad():
for i, (src, tgt) in enumerate(self.test_dataloader):
out = self.forward_net(src, tgt)
# out = torch.max(out, 2).indices
loss = loss_f.forward(out.reshape(-1, voacb_size), tgt[:, 1:].flatten())
loss_a += loss.item()
self.progress.print(
f"Validation: epoch={self.epoch} avg loss={loss_a/len(self.test_dataloader):.4f}"
)
def training(self, train_epoch_nums: int = 36):
self.progress.start()
training_all = self.progress.add_task(
description=f"epoch={self.epoch} lr={self.optimizer_scheduler.get_lr()}",
total=train_epoch_nums,
)
for i in range(train_epoch_nums):
self.progress.update(
training_all,
advance=1,
description=f"epoch={self.epoch} lr={self.optimizer_scheduler.get_lr()}",
)
self.train_epoch()
self.evaluation()
self.epoch += 1
self.save_checkpoint()
print(self.generate(4, "绿蔓如藤不用栽"))
def main():
model = PoetryGen()
# print(model.generate())
while (s := input(">")) != "exit":
if s.find("/") != -1:
print(model.generate_by_start(s, "绿蔓如藤不用栽"))
else:
print(model.generate(4, s or "床前明月光"))
# model.training(256)
if __name__ == "__main__":
main()