-
Notifications
You must be signed in to change notification settings - Fork 3
/
trainer.py
213 lines (172 loc) · 8.17 KB
/
trainer.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
import math
import sys, os
import os.path as osp
import time
import pickle
import numpy as np
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.nn.functional as F
from utils.exp_utils import create_exp_dir
from utils.utils import *
from utils.cp_data_utils import *
from utils.cp_function_utils import *
from models import cord_cpd
class Trainer(object):
def __init__(self, args):
super(Trainer, self).__init__()
self.args = args
self.args.cuda = not args.no_cuda and torch.cuda.is_available()
self.device = torch.device("cuda") if self.args.cuda else torch.device("cpu")
self.args.factor = not args.no_factor
self.exp_dir = args.exp_dir
seed = args.seed
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
self.logging = create_exp_dir(args.exp_dir)
meta_file_name = osp.join(args.exp_dir, "meta.txt")
meta_file = open(meta_file_name, "w")
meta_file.write(str(args))
meta_file.close()
def load_data(self):
args = self.args
if self.data_type == "sim":
self.train_loader, self.valid_loader, self.test_loader = load_cp_data(
args.data_path, args.batch_size, args.suffix, args.data_norm)
else:
self.train_loader, self.valid_loader, self.test_loader = load_cp_real_data(
args.data_path, args.batch_size)
off_diag = np.ones([args.num_atoms, args.num_atoms]) - np.eye(args.num_atoms)
rel_rec = np.array(encode_onehot(np.where(off_diag)[1]), dtype=np.float32)
rel_send = np.array(encode_onehot(np.where(off_diag)[0]), dtype=np.float32)
self.rel_rec = torch.FloatTensor(rel_rec).to(self.device)
self.rel_send = torch.FloatTensor(rel_send).to(self.device)
def set_model(self):
self.model = cord_cpd.Model(self.args)
self.model.to_device(self.device)
if self.args.load:
self.logging("loading model from {}".format(self.args.exp_dir))
self.model.load(self.exp_dir)
def train(self):
# Train model
st = time.time()
best_val_loss = np.inf
best_acc_val = 0
best_epoch = 0
for epoch in range(self.args.epochs):
t = time.time()
mse_train, delta_train, acc_train = self.train_one_epoch()
log_str_train = "Epoch: {:4d}, mse_train: {:.4f}, delta_train: {:.4f}, " \
"acc_train: {:.4f}, epoch time: {:.2f}s".format(
epoch, mse_train, delta_train,
acc_train, time.time() - t
)
log_str_eval = ""
if (epoch+1) % self.args.eval_epoch == 0:
mse_val, delta_val, acc_val, avg_roc, avg_dist, avg_tri = self.evaluate()
log_str_eval = "|| mse_val: {:.4f}, delta_val: {:.4f}, acc_val: {:.4f}, " \
"roc: {:.4f}, dist: {:.4f}, tri: {:.4f}, total time: {:.2f}s ||".format(
mse_val, delta_val, acc_val,
avg_roc, avg_dist, avg_tri, time.time()-st
)
if mse_val < best_val_loss:
best_val_loss = mse_val
best_epoch = epoch
self.model.save(self.exp_dir)
self.logging("save best model at epoch : {}".format(best_epoch))
if acc_val > best_acc_val:
best_acc_val = acc_val
self.model.save(self.exp_dir, "acc_model.t7")
self.logging("save acc model at epoch : {}".format(epoch))
self.logging(log_str_train + log_str_eval)
self.logging("Optimization Finished!")
self.logging("Best Epoch: {:04d}".format(best_epoch))
def train_one_epoch(self):
acc_train = []
mse_train = []
delta_train = []
self.model.set_train()
for batch_idx, (data, relations, cpd) in enumerate(self.train_loader):
data, relations = data.to(self.device), relations.to(self.device)
# data [batch_size, num_atoms, num_timesteps, num_dims]
data = data[:, :, :self.args.timesteps, :]
self.model.optimizer.zero_grad()
logits = self.model.encode(data, self.rel_rec, self.rel_send)
# loss_delta = 10000 * ((logits[:, :-1] - logits[:, 1:]) ** 2).mean()
# logits [batch, timestep, edge, relation]
sub_logits = logits[:, 5:-5]
loss_delta = 100 * ((sub_logits[:, :-1] - sub_logits[:, 1:]) ** 2).mean()
edges = F.gumbel_softmax(logits, tau=self.args.temp, hard=self.args.hard)
# prob = F.softmax(logits, -1)
output = self.model.decode(data, edges, self.rel_rec, self.rel_send)
target = data[:, :, 1:, :]
target = target[:, :, self.args.begin_steps:, :]
output = output[:, :, self.args.begin_steps:, :]
loss_mse = F.mse_loss(output, target) / (2 * self.args.var) * 400
loss = loss_mse + loss_delta
if self.data_type == "sim":
acc = edge_accuracy(logits, relations, begin_steps=5, end_steps=-5)
acc_train.append(acc)
else:
acc_train.append(np.nan)
loss.backward()
self.model.optimizer.step()
mse_train.append(loss_mse.item())
delta_train.append(loss_delta.item())
self.model.scheduler.step()
return np.mean(mse_train), np.mean(delta_train), np.mean(acc_train)
@torch.no_grad()
def evaluate(self):
acc_val = []
mse_val = []
delta_val = []
self.model.set_eval()
probs = []
cpds = []
recons = []
origs = []
for batch_idx, (data, relations, cpd) in enumerate(self.valid_loader):
data, relations = data.to(self.device), relations.to(self.device)
data = data[:, :, :self.args.timesteps, :]
logits = self.model.encode(data, self.rel_rec, self.rel_send)
sub_logits = logits[:, 5:-5]
loss_delta = 100 * ((sub_logits[:, :-1] - sub_logits[:, 1:]) ** 2).mean()
edges = F.gumbel_softmax(logits, tau=self.args.temp, hard=True)
prob = F.softmax(logits, -1)
probs.append(prob)
cpds.extend(cpd)
output = self.model.decode(data, edges, self.rel_rec, self.rel_send)
target = data[:, :, 1:, :]
target = target[:, :, self.args.begin_steps:, :]
output = output[:, :, self.args.begin_steps:, :]
loss_mse = F.mse_loss(output, target) / (2 * self.args.var) * 400
if self.data_type == 'sim':
acc = edge_accuracy(logits, relations, begin_steps=5, end_steps=-5)
acc_val.append(acc)
else:
origs.append(data.transpose(1, 2).contiguous().detach().cpu().numpy())
# validation output uses teacher forcing
recon = self.model.decoder.forward_reconstruct(data, edges, self.rel_rec, self.rel_send)
recons.append(recon.detach().cpu().numpy())
acc_val.append(np.nan)
mse_val.append(loss_mse.item())
delta_val.append(loss_delta.item())
probs = torch.cat(probs).detach().cpu().numpy()
cpds = np.array(cpds)
avg_roc, avg_dist, avg_tri = cpd_metrics(probs, cpds)
if self.report_combine:
recons = np.concatenate(recons)
origs = np.concatenate(origs)
type1_score = mse_anomaly(recons, origs, step=5)
type2_score = cal_cp_from_output(probs)
combined = anomaly_combined_score(type1_score, type2_score)
self.logging("-"*30)
self.logging("relation score: {}".format(cpd_metrics(type2_score, cpds, anomaly_input=True)))
self.logging("mse score: {}".format(cpd_metrics(type1_score, cpds, anomaly_input=True)))
self.logging("combined score: {}".format(cpd_metrics(combined, cpds, anomaly_input=True)))
self.logging("-" * 30)
self.model.set_train()
return np.mean(mse_val), np.mean(delta_val), np.mean(acc_val), avg_roc, avg_dist, avg_tri