-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
263 lines (230 loc) · 10.4 KB
/
train.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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
import argparse
import os
from utils import RoundSTE
import sys
from cvae import CVAE
import torch
import torch.nn.functional as F
import torch_geometric
from torch.optim import Adam
from torch.optim.lr_scheduler import LambdaLR
from torch.utils.data import DataLoader, random_split
import data_utils
import diffusion
from decoder import SolutionDecoder
from cisp import CISP
from tqdm import tqdm
def lr_lambda(epoch):
return 0.9 ** ((epoch + 1) // 15)
def set_model_to_mode(model, mode):
if mode == 'eval':
if isinstance(model, list):
for m in model:
m.eval()
else:
model.eval()
else:
if isinstance(model, list):
for m in model:
m.train()
else:
model.train()
def Loss_CV(mip, sol_per, ptype):
# lamda * sum(max( ax - b , 0 )); lamda = n_vars if type != IS else 0
lamda = mip.n_vars[0]
sol = sol_per
if ptype == "IS":
return torch.tensor(0)
else:
A, b, _ = mip.getCoff()
result = torch.sparse.mm(A, sol.unsqueeze(1))
Ax_minus_b = result.squeeze(1) - b
max_violations = torch.clamp(Ax_minus_b, min=0).mean()
loss = lamda * max_violations
return loss
def forward_by_vae(mip, x, model, checkpoint=None):
if checkpoint is not None:
model.load_state_dict(checkpoint['ddpm_state_dict'])
n_int_var = mip.n_int_vars
zi, _ = vae.encode_mip(mip, n_int_var)
zx, key = vae.encode_solution(x, n_int_var)
zx_start, loss_ddpm = model(zx, zi, key)
sols = vae.decoder(zi, zx_start, key)
loss_CV = Loss_CV(mip, sols, args.type)
return loss_ddpm, loss_CV, sols
def forward_by_cisp(mip, x, model, checkpoint=None):
ddpm, decoder = model
if checkpoint is not None:
ddpm.load_state_dict(checkpoint['ddpm_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
n_int_var = mip.n_int_vars
zi, _ = cisp.encode_mip(mip, n_int_var)
zx, key = cisp.encode_solution(x, n_int_var)
zx_start, loss_ddpm = ddpm(zx, zi, key)
sols = decoder(zi, zx_start, key)
loss_decoder = F.binary_cross_entropy(sols, x.float())
# sols_round = RoundSTE.apply(sols)
loss_CV = Loss_CV(mip, sols, args.type)
return loss_ddpm, loss_CV, loss_decoder, sols
def train_one_epoch(model, optimizer, scheduler, data_loader, device, epoch, checkpoint=None, tb_writer=None):
if checkpoint is not None:
check = torch.load(checkpoint)
epoch = check['epoch']
optimizer.load_state_dict(check['optimizer_state_dict'])
scheduler.load_state_dict(check['scheduler_state_dict'])
else:
check = None
set_model_to_mode(model, 'train')
mean_loss = torch.zeros(1).to(device)
mean_loss_ddpm = torch.zeros(1).to(device)
mean_loss_CV = torch.zeros(1).to(device)
accumulation_steps = 8
data_loader = tqdm(data_loader)
for iteration, mip in enumerate(data_loader):
x = mip.sols
if args.vae:
loss_ddpm, loss_CV, sols = forward_by_vae(mip, x, model, check)
loss = loss_ddpm + loss_CV
else:
loss_ddpm, loss_CV, loss_decoder, sols = forward_by_cisp(mip, x, model, check)
# 必须decoder * 10
loss = loss_ddpm + loss_CV + loss_decoder * 20
loss.backward()
mean_loss_ddpm = (mean_loss_ddpm * iteration + loss_ddpm.detach()) / (iteration + 1)
mean_loss_CV = (mean_loss_CV * iteration + loss_CV.detach()) / (iteration + 1)
mean_loss = (mean_loss * iteration + loss.detach()) / (iteration + 1)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss)
sys.exit(1)
if (iteration + 1) % accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
data_loader.desc = "[epoch {}] loss {} loss_CV {} loss_ddpm {}".format(epoch, round(mean_loss.item(), 4),
round(mean_loss_CV.item(), 2),
round(mean_loss_ddpm.item(), 4))
if iteration == len(data_loader) - 1:
optimizer.step()
optimizer.zero_grad()
data_loader.desc = "[epoch {}] mean loss {}".format(epoch, round(mean_loss.item(), 4))
if tb_writer is not None:
tags = ["train_loss", "learning_rate"]
# tensorboard可视化
for tag, value in zip(tags, [mean_loss.item(), optimizer.param_groups[0]["lr"]]):
tb_writer.add_scalars('Train %s' % tag, value, iteration)
scheduler.step()
return mean_loss.item(), mean_loss_CV.item()
def evaluate(model, data_loader, epoch):
# set_model_to_mode(model, 'eval')
total_val_loss = 0
total_ddpm_loss = 0
total_CV_loss = 0
total_obj = 0
fea = 0
with torch.no_grad():
for iteration, mip in enumerate(data_loader):
x = mip.sols
if args.vae:
loss_ddpm, loss_CV, sols = forward_by_vae(mip, x, model)
loss = loss_ddpm + loss_CV
else:
loss_ddpm, loss_CV, loss_decoder, sols = forward_by_cisp(mip, x, model)
loss = loss_ddpm + loss_CV + loss_decoder
sols_round = sols.round()
A, b, c = mip.getCoff()
obj = sols_round.squeeze() @ c
violates = torch.max((A @ sols_round).squeeze() - b,
torch.tensor(0)).mean()
if violates <= 0:
fea += args.batch
total_val_loss += loss.item()
total_ddpm_loss += loss_ddpm.item()
total_CV_loss += loss_CV.item()
total_obj += obj.item()
avg_val_loss = total_val_loss / (iteration + 1)
avg_ddpm_loss = total_ddpm_loss / (iteration + 1)
avg_CV_loss = total_CV_loss / (iteration + 1)
avg_obj = total_obj / (iteration + 1)
logger.logger.info(
f'Epoch: {epoch}, Validation Loss: {avg_val_loss} Loss_ddpm: {avg_ddpm_loss} Loss_CV:{avg_CV_loss}')
logger.logger.info(f'Epoch: {epoch}, Validation fea: {fea} / 100, obj:{avg_obj}')
return avg_val_loss
def main(args, logger):
# checkpoint = f'./model/{args.type}/best_checkpoint_vae_False.pth'
# modelPath = os.path.join(f'./model/{args.type}', checkpoint)
# if not os.path.isdir(modelPath):
# checkpoint = None
# logger.logger.info('No such checkpoint!')
ddpm = diffusion.DDPMTrainer(attn_dim=128, n_heads=4, n_layers=1, device=device,
parameterization=f'{args.p}')
ddpm.to(device)
if not args.vae:
decoder = SolutionDecoder(attn_dim=128, n_heads=4, n_layers=2, attn_mask=None)
decoder.to(device)
optimizer = Adam([
{'params': ddpm.parameters(), 'lr': 0.0008},
{'params': decoder.parameters(), 'lr': 0.0005}
])
model = [ddpm, decoder]
else:
optimizer = Adam(ddpm.parameters(), 0.0008)
model = ddpm
epochs = 100
scheduler = LambdaLR(optimizer, lr_lambda)
logger.logger.info(f'start training {args.type}...... vae {args.vae}\n')
best_val_loss = float('inf')
optimizer.zero_grad()
for epoch in range(epochs):
mean_loss, mean_loss_CV = train_one_epoch(model, optimizer, scheduler, train_loader, device, epoch,
checkpoint=None)
logger.logger.info('%d epoch train mean loss: %.4f CV_loss: %.4f\n' % (epoch, mean_loss, mean_loss_CV))
val_loss = evaluate(model, val_loader, epoch)
if epoch == 30:
print('stop')
if epoch + 1 >= args.save_epoch and (epoch + 1) % args.save_epoch == 0:
checkpoint = {
'ddpm_state_dict': ddpm.state_dict(), # *模型参数
'decoder_state_dict': decoder.state_dict() if args.vae is not True else None,
'optimizer_state_dict': optimizer.state_dict(), # *优化器参数
'scheduler_state_dict': scheduler.state_dict(), # *scheduler
'epoch': epoch
}
torch.save(checkpoint, os.path.join(path, f'checkpoint-%d-{args.vae}.pth' % epoch))
logger.logger.info('save model %d successed......\n' % epoch)
if epoch + 1 >= args.save_epoch and val_loss < best_val_loss:
best_val_loss = val_loss
logger.logger.info('best model in %d epoch, validation loss: %.6f \n' % (epoch, val_loss))
checkpoint = {
'ddpm_state_dict': ddpm.state_dict(), # *模型参数
'decoder_state_dict': decoder.state_dict() if args.vae is not True else None,
'optimizer_state_dict': optimizer.state_dict(), # *优化器参数
'scheduler_state_dict': scheduler.state_dict(), # *scheduler
'epoch': epoch,
}
torch.save(checkpoint, os.path.join(path, f'best_checkpoint_vae_{args.vae}.pth'))
logger.logger.info('save best model successed......\n')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--save_epoch', type=float, default=10)
parser.add_argument('--log_dir', type=str, default='./Logs/summary')
parser.add_argument('--type', type=str, default="CA")
parser.add_argument('--vae', type=bool, default=False)
parser.add_argument('--batch', type=int, default=4)
parser.add_argument('--p', type=str, default='x0', help='whether eps or x0 the ddpm predict')
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
args = parser.parse_args()
graphSet = data_utils.GraphDataset(args.type)
total_size = len(graphSet)
train_size = int(total_size * 0.9)
val_size = total_size - train_size
train_set, val_set = random_split(graphSet, [train_size, val_size], generator=torch.Generator().manual_seed(2024))
train_loader = torch_geometric.data.DataLoader(train_set, batch_size=args.batch, shuffle=True)
val_loader = torch_geometric.data.DataLoader(val_set, batch_size=args.batch, shuffle=False)
path = f'./model/{args.type}'
cisp = CISP()
cisp.to(device)
cisp.load_state_dict(torch.load(f'./model/{args.type}/cisp_pre/best_checkpoint.pth')['model_state_dict'])
vae = CVAE(embedding=True)
vae.to(device)
vae.load_state_dict(torch.load(f'./model/{args.type}/cvae_embedding_pre/best_checkpoint.pth')['model_state_dict'])
logger = data_utils.Logger(args)
main(args, logger)