trainer.py 9.46 KiB
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
import torch
import os
import time
import argparse
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import math
from model import *
from tqdm import *
from util import KpiReader
from logger import Logger
class Trainer(object):
def __init__(self, model, train, trainloader, log_path='log_trainer',
log_file='loss', epochs=50, batch_size=64, learning_rate=0.001,
checkpoints='kpi_model.path', checkpoints_interval = 10, device=torch.device('cuda:0')):
self.trainloader = trainloader
self.train = train
self.log_path = log_path
self.log_file = log_file
self.start_epoch = 0
self.epochs = epochs
self.device = device
self.batch_size = batch_size
self.model = model
self.model.to(device)
self.learning_rate = learning_rate
self.checkpoints = checkpoints
self.checkpoints_interval = checkpoints_interval
self.optimizer = optim.Adam(self.model.parameters(), self.learning_rate)
self.epoch_losses = []
self.loss = {}
self.logger = Logger(self.log_path, self.log_file)
def save_checkpoint(self, epoch):
torch.save({'epoch': epoch + 1,
'state_dict': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'losses': self.epoch_losses},
self.checkpoints + '_epochs{}.pth'.format(epoch+1))
def load_checkpoint(self, start_ep):
try:
print ("Loading Chechpoint from ' {} '".format(self.checkpoints+'_epochs{}.pth'.format(start_ep)))
checkpoint = torch.load(self.checkpoints+'_epochs{}.pth'.format(start_ep))
self.start_epoch = checkpoint['epoch']
self.model.load_state_dict(checkpoint['state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.epoch_losses = checkpoint['losses']
print ("Resuming Training From Epoch {}".format(self.start_epoch))
except:
print ("No Checkpoint Exists At '{}', Starting Fresh Training".format(self.checkpoints))
self.start_epoch = 0
def loss_fn(self, original_seq, recon_seq_mu, recon_seq_logvar, s_mean,
s_logvar, d_post_mean, d_post_logvar, d_prior_mean, d_prior_logvar):
batch_size = original_seq.size(0)
# See https://arxiv.org/pdf/1606.05908.pdf, Page 9, Section 2.2 for details.
# N(x|mu,var)
# = log{1/(sqrt(2*pi)*var)exp{-(x-mu)^2/(2*var^2)}}
# = -0.5*{log(2*pi)+2*log(var)+[(x-mu)/exp{log(var)}]^2}
loglikelihood = -0.5 * torch.sum(torch.pow(((original_seq.float()-recon_seq_mu.float())/torch.exp(recon_seq_logvar.float())), 2)
+ 2 * recon_seq_logvar.float()
+ np.log(np.pi*2))
# See https://arxiv.org/pdf/1606.05908.pdf, Page 9, Section 2.2, Equation (7) for details.
kld_s = -0.5 * torch.sum(1 + s_logvar - torch.pow(s_mean, 2) - torch.exp(s_logvar))
# See https://arxiv.org/pdf/1606.05908.pdf, Page 9, Section 2.2, Equation (6) for details.
d_post_var = torch.exp(d_post_logvar)
d_prior_var = torch.exp(d_prior_logvar)
kld_d = 0.5 * torch.sum(d_prior_logvar - d_post_logvar
+ ((d_post_var + torch.pow(d_post_mean - d_prior_mean, 2)) / d_prior_var)
- 1)
return (-loglikelihood + kld_s + kld_d)/batch_size, loglikelihood/batch_size, kld_s/batch_size, kld_d/batch_size
def train_model(self):
self.model.train()
for epoch in range(self.start_epoch, self.epochs):
losses = []
llhs = []
kld_ss = []
kld_ds = []
print ("Running Epoch : {}".format(epoch+1))
for i, dataitem in tqdm(enumerate(self.trainloader,1)):
_,_,data = dataitem
data = data.to(self.device)
self.optimizer.zero_grad()
s_mean, s_logvar, s, d_post_mean, d_post_logvar, d, d_prior_mean, d_prior_logvar, recon_x_mu, recon_x_logvar = self.model(data)
loss, llh, kld_s, kld_d = self.loss_fn(data, recon_x_mu, recon_x_logvar, s_mean, s_logvar,
d_post_mean, d_post_logvar, d_prior_mean, d_prior_logvar)
loss.backward()
self.optimizer.step()
losses.append(loss.item())
llhs.append(llh.item())
kld_ss.append(kld_s.item())
kld_ds.append(kld_d.item())
meanloss = np.mean(losses)
meanllh = np.mean(llhs)
means = np.mean(kld_ss)
meand = np.mean(kld_ds)
self.epoch_losses.append(meanloss)
print ("Epoch {} : Average Loss: {} Loglikelihood: {} KL of s: {} KL of d: {}".format(epoch+1, meanloss, meanllh, means, meand))
self.loss['Epoch'] = epoch+1
self.loss['Avg_loss'] = meanloss
self.loss['Llh'] = meanllh
self.loss['KL_s'] = means
self.loss['KL_d'] = meand
self.logger.log_trainer(epoch+1, self.loss)
if (self.checkpoints_interval > 0
and (epoch+1) % self.checkpoints_interval == 0):
self.save_checkpoint(epoch)
print ("Training is complete!")
def main():
parser = argparse.ArgumentParser()
# GPU option
parser.add_argument('--gpu_id', type=int, default=0)
# Dataset options
parser.add_argument('--dataset_path', type=str, default='')
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--num_workers', type=int, default=4)
parser.add_argument('--T', type=int, default=20)
parser.add_argument('--win_size', type=int, default=36)
parser.add_argument('--l', type=int, default=10)
parser.add_argument('--n', type=int, default=24)
# Model options
parser.add_argument('--s_dims', type=int, default=8)
parser.add_argument('--d_dims', type=int, default=10)
parser.add_argument('--conv_dims', type=int, default=100)
parser.add_argument('--hidden_dims', type=int, default=40)
parser.add_argument('--enc_dec', type=str, default='CNN')
# Training options
parser.add_argument('--learning_rate', type=float, default=0.0002)
parser.add_argument('--epochs', type=int, default=50)
parser.add_argument('--start_epoch', type=int, default=0)
parser.add_argument('--checkpoints_path', type=str, default='model')
parser.add_argument('--checkpoints_file', type=str, default='')
parser.add_argument('--checkpoints_interval', type=int, default=10)
parser.add_argument('--log_path', type=str, default='log_trainer')
parser.add_argument('--log_file', type=str, default='')
args = parser.parse_args()
# Set up GPU
if torch.cuda.is_available() and args.gpu_id >= 0:
device = torch.device('cuda:%d' % args.gpu_id)
else:
device = torch.device('cpu')
if not os.path.exists(args.dataset_path):
raise ValueError('Unknown dataset path: {}'.format(args.dataset_path))
if not os.path.exists(args.log_path):
os.makedirs(args.log_path)
if not os.path.exists(args.checkpoints_path):
os.makedirs(args.checkpoints_path)
if args.checkpoints_file == '':
args.checkpoints_file = 'sdim{}_ddim{}_cdim{}_hdim{}_winsize{}_T{}_l{}'.format(
args.s_dims,
args.d_dims,
args.conv_dims,
args.hidden_dims,
args.win_size,
args.T,args.l)
if args.log_file == '':
args.log_file = 'sdim{}_ddim{}_cdim{}_hdim{}_winsize{}_T{}_l{}_loss'.format(
args.s_dims,
args.d_dims,
args.conv_dims,
args.hidden_dims,
args.win_size,
args.T,args.l)
kpi_value_train = KpiReader(args.dataset_path)
train_loader = torch.utils.data.DataLoader(kpi_value_train,
batch_size = args.batch_size,
shuffle = True,
num_workers = args.num_workers)
sdfvae = SDFVAE(s_dim = args.s_dims,
d_dim = args.d_dims,
conv_dim = args.conv_dims,
hidden_dim = args.hidden_dims,
T = args.T,
w = args.win_size,
n = args.n,
enc_dec = args.enc_dec,
device = device)
trainer = Trainer(sdfvae, kpi_value_train, train_loader,
log_path = args.log_path,
log_file = args.log_file,
batch_size = args.batch_size,
epochs = args.epochs,
learning_rate = args.learning_rate,
checkpoints = os.path.join(args.checkpoints_path,args.checkpoints_file),
checkpoints_interval = args.checkpoints_interval,
device = device)
trainer.load_checkpoint(args.start_epoch)
trainer.train_model()
if __name__ == '__main__':
main()