-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainner.py
190 lines (162 loc) · 8.43 KB
/
trainner.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
import time
import torch
import numpy as np
from tqdm import tqdm
import random
from torch_geometric.utils import negative_sampling
from sklearn.metrics import roc_auc_score, average_precision_score
import warnings
warnings.filterwarnings("ignore")
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
class Trainer(object):
def __init__(self, args, model, data, writer=None, **kwargs):
self.args = args
self.data = data
self.model = model
self.writer = writer
self.len = len(data['train']['edge_index_list'])
self.len_train = self.len - args.testlength - args.vallength
self.len_val = args.vallength
self.len_test = args.testlength
x = data['x'].to(args.device)
self.x = [x for _ in range(self.len)] if len(x.shape) <= 2 else x
setup_seed(args.seed)
print('total length: {}, test length: {}'.format(
self.len, args.testlength))
def run(self):
args = self.args
max_auc = 0
max_test_auc = 0
max_train_auc = 0
min_epoch = args.min_epoch
max_patience = args.patience
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
t_total0 = time.time()
with tqdm(range(1, args.max_epoch + 1)) as bar:
for epoch in bar:
t0 = time.time()
average_epoch_loss, average_train_auc, average_val_auc, average_test_auc = self.train(epoch, self.data['train'])
# update the best results.
if average_val_auc > max_auc:
max_auc = average_val_auc
max_test_auc = average_test_auc
max_train_auc = average_train_auc
# ood results
test_results = self.test(epoch, self.data['test'])
patience = 0
else:
patience += 1
if epoch > min_epoch and patience > max_patience:
break
if epoch == 1 or epoch % self.args.log_interval == 0:
print(
"Epoch:{}, Loss: {:.4f}, Time: {:.3f}".format(epoch, average_epoch_loss, time.time() - t0))
print(
f"(IID) Current: Epoch:{epoch}, Train AUC :{average_train_auc:.4f}, Val AUC: {average_val_auc:.4f}, Test AUC: {average_test_auc:.4f}"
)
print(
f"(IID) Best_Test: Epoch:{test_results[0]}, Train AUC:{max_train_auc:.4f}, Val AUC: {max_auc:.4f}, Test AUC: {max_test_auc:.4f}"
)
print(
f"(OOD) Current_Test: Epoch:{test_results[0]}, Train AUC:{test_results[1]:.4f}, Val AUC: {test_results[2]:.4f}, Test AUC: {test_results[3]:.4f}"
)
epoch_time = (time.time() - t_total0) / (epoch - 1)
return max_train_auc, max_auc, max_test_auc, test_results, epoch_time, epoch, average_train_auc, average_val_auc, average_test_auc
def train(self, epoch, data):
self.model.train()
optimizer = self.optimizer
causal_embedding_list, conf_embedding_list, env_kl_loss = \
self.model(data['edge_index_list'], self.x, self.len_train)
conf_embedding_list = [emb.detach() for emb in conf_embedding_list]
causal_loss = torch.tensor([]).to(self.args.device)
env_loss = torch.tensor([]).to(self.args.device)
for t in range(self.len_train-1):
causal_embedding = causal_embedding_list[:, t, :].squeeze() #[N, F]
if self.args.dataset == 'act':
pos_edge_index = data['pedges'][t + 1].long().to(self.args.device)
neg_edge_index = negative_sampling(pos_edge_index, self.args.num_nodes, num_neg_samples= pos_edge_index.shape[1])
else:
pos_edge_index, neg_edge_index = data['pedges'][t+1].long().to(self.args.device),data['nedges'][t+1].long().to(self.args.device)
causal_pos_pred, causal_neg_pred = self.cal_pred(causal_embedding, pos_edge_index, neg_edge_index,
self.model.cs_decoder)
causal_loss = torch.cat([causal_loss, self.loss(causal_pos_pred, causal_neg_pred).unsqueeze(0)])
# causal intervention
for times in range(self.args.intervention_times):
n_s = np.random.randint(self.len_train)
conf_z = conf_embedding_list[n_s]
conf_pos_pred, conf_neg_pred = self.cal_pred(conf_z, pos_edge_index, neg_edge_index,
self.model.ss_decoder)
s1 = np.random.randint(len(conf_pos_pred))
conf_pos_pred_s = torch.sigmoid(conf_pos_pred[s1]).detach()
conf_neg_pred_s = torch.sigmoid(conf_neg_pred[s1]).detach()
conf_pos = conf_pos_pred_s * causal_pos_pred
conf_neg = conf_neg_pred_s * causal_neg_pred
env_loss = torch.cat([env_loss, self.loss(conf_pos, conf_neg).unsqueeze(0)])
env_mean = env_loss.mean()
env_var = torch.var(env_loss * self.args.intervention_times)
penalty = env_mean + env_var
la = self.args.weight1
if epoch < self.args.warm_epoch:
la = 0
loss = torch.mean(causal_loss) + la * penalty + self.args.weight2 *torch.mean(env_kl_loss)
# loss = torch.mean(causal_loss)
optimizer.zero_grad()
loss.backward()
optimizer.step()
average_epoch_loss = loss.item()
# train_test(iid data)
self.model.eval()
train_auc_list, val_auc_list, test_auc_list = [], [], []
causal_embedding_list, _, _ = \
self.model(data['edge_index_list'], self.x, self.len)
for t in range(self.len - 1):
z_mix = causal_embedding_list[:, t, :].squeeze() #[N, F]
pos_edge, neg_edge = data['pedges'][t+1].long().to(self.args.device),data['nedges'][t+1].long().to(self.args.device)
auc_mix, ap_mix = self.predict(z_mix, pos_edge, neg_edge, self.model.cs_decoder)
if t < self.len_train - 1:
train_auc_list.append(auc_mix)
elif t < self.len_train + self.len_val - 1:
val_auc_list.append(auc_mix)
else:
test_auc_list.append(auc_mix)
return average_epoch_loss, np.mean(train_auc_list), np.mean(val_auc_list), np.mean(test_auc_list)
def test(self, epoch, data):
self.model.eval()
causal_embedding_list, _, _ = self.model(data['edge_index_list'], self.x, self.len)
train_auc_list, val_auc_list, test_auc_list = [], [], []
for t in range(self.len - 1):
z_mix = causal_embedding_list[:, t, :].squeeze() #[N, F]
pos_edge, neg_edge = data['pedges'][t + 1].long().to(self.args.device), data['nedges'][t + 1].long().to(
self.args.device)
if neg_edge.shape[1]==0:
continue
auc_mix, ap_mix = self.predict(z_mix, pos_edge, neg_edge, self.model.cs_decoder)
if t < self.len_train - 1:
train_auc_list.append(auc_mix)
elif t < self.len_train + self.len_val - 1:
val_auc_list.append(auc_mix)
else:
test_auc_list.append(auc_mix)
return epoch, np.mean(train_auc_list), np.mean(val_auc_list), np.mean(test_auc_list)
def cal_pred(self, z, edge_index_pos, edge_index_neg, decoder):
pos_pred = decoder(z, edge_index_pos)
neg_pred = decoder(z, edge_index_neg)
return pos_pred, neg_pred
def loss(self, pos_pred, neg_pred):
pos_loss = -torch.log(pos_pred + 1e-9).mean()
neg_loss = -torch.log(1 - neg_pred + 1e-9).mean()
return pos_loss + neg_loss
def predict(self, z, pos_edge_index, neg_edge_index, decoder):
pos_y = z.new_ones(pos_edge_index.size(1)).to(z.device)
neg_y = z.new_zeros(neg_edge_index.size(1)).to(z.device)
y = torch.cat([pos_y, neg_y], dim=0)
pos_pred = decoder(z, pos_edge_index)
neg_pred = decoder(z, neg_edge_index)
pred = torch.cat([pos_pred, neg_pred], dim=0)
y, pred = y.detach().cpu().numpy(), pred.detach().cpu().numpy()
return roc_auc_score(y, pred), average_precision_score(y, pred)