-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrain.py
157 lines (130 loc) · 6.85 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
import os
import sys
import time
import numpy as np
import pickle
import scipy.sparse as sp
import logging
import argparse
import torch
import torch.nn.functional as F
from sklearn.metrics import roc_auc_score
from model import Model
from preprocess import normalize_sym, normalize_row, sparse_mx_to_torch_sparse_tensor
from arch import archs
parser = argparse.ArgumentParser()
parser.add_argument('--lr', type=float, default=0.01, help='learning rate')
parser.add_argument('--wd', type=float, default=0.001, help='weight decay')
parser.add_argument('--n_hid', type=int, default=64, help='hidden dimension')
parser.add_argument('--dataset', type=str, default='Yelp')
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--epochs', type=int, default=200, help='number of training epochs')
parser.add_argument('--dropout', type=float, default=0.2)
parser.add_argument('--seed', type=int, default=1)
args = parser.parse_args()
prefix = "lr" + str(args.lr) + "_wd" + str(args.wd) + "_h" + str(args.n_hid) + \
"_drop" + str(args.dropout) + "_epoch" + str(args.epochs) + "_cuda" + str(args.gpu)
logdir = os.path.join("log/eval", args.dataset)
if not os.path.exists(logdir):
os.makedirs(logdir)
log_format = '%(message)s'
logging.basicConfig(stream=sys.stdout, level=logging.INFO, format=log_format)
fh = logging.FileHandler(os.path.join(logdir, prefix + ".txt"))
fh.setFormatter(logging.Formatter(log_format))
logging.getLogger().addHandler(fh)
def main():
torch.cuda.set_device(args.gpu)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
steps_s = [len(meta) for meta in archs[args.dataset]["source"][0]]
steps_t = [len(meta) for meta in archs[args.dataset]["target"][0]]
#print(steps_s, steps_t)
datadir = "preprocessed"
prefix = os.path.join(datadir, args.dataset)
#* load data
node_types = np.load(os.path.join(prefix, "node_types.npy"))
num_node_types = node_types.max() + 1
node_types = torch.from_numpy(node_types).cuda()
adjs_offset = pickle.load(open(os.path.join(prefix, "adjs_offset.pkl"), "rb"))
adjs_pt = []
if '0' in adjs_offset:
adjs_pt.append(sparse_mx_to_torch_sparse_tensor(normalize_sym(adjs_offset['0'] + sp.eye(adjs_offset['0'].shape[0], dtype=np.float32))).cuda())
for i in range(1, int(max(adjs_offset.keys())) + 1):
adjs_pt.append(sparse_mx_to_torch_sparse_tensor(normalize_row(adjs_offset[str(i)] + sp.eye(adjs_offset[str(i)].shape[0], dtype=np.float32))).cuda())
adjs_pt.append(sparse_mx_to_torch_sparse_tensor(normalize_row(adjs_offset[str(i)].T + sp.eye(adjs_offset[str(i)].shape[0], dtype=np.float32))).cuda())
adjs_pt.append(sparse_mx_to_torch_sparse_tensor(sp.eye(adjs_offset['1'].shape[0], dtype=np.float32).tocoo()).cuda())
adjs_pt.append(torch.sparse.FloatTensor(size=adjs_offset['1'].shape).cuda())
print("Loading {} adjs...".format(len(adjs_pt)))
#* load labels
pos = np.load(os.path.join(prefix, "pos_pairs_offset.npz"))
pos_train = pos['train']
pos_val = pos['val']
pos_test = pos['test']
neg = np.load(os.path.join(prefix, "neg_pairs_offset.npz"))
neg_train = neg['train']
neg_val = neg['val']
neg_test = neg['test']
#* one-hot IDs as input features
in_dims = []
node_feats = []
for k in range(num_node_types):
in_dims.append((node_types == k).sum().item())
i = torch.stack((torch.arange(in_dims[-1], dtype=torch.long), torch.arange(in_dims[-1], dtype=torch.long)))
v = torch.ones(in_dims[-1])
node_feats.append(torch.sparse.FloatTensor(i, v, torch.Size([in_dims[-1], in_dims[-1]])).cuda())
assert(len(in_dims) == len(node_feats))
model_s = Model(in_dims, args.n_hid, steps_s, dropout = args.dropout).cuda()
model_t = Model(in_dims, args.n_hid, steps_t, dropout = args.dropout).cuda()
optimizer = torch.optim.Adam(
list(model_s.parameters()) + list(model_t.parameters()),
lr=args.lr,
weight_decay=args.wd
)
best_val = None
final = None
anchor = None
for epoch in range(args.epochs):
train_loss = train(node_feats, node_types, adjs_pt, pos_train, neg_train, model_s, model_t, optimizer)
val_loss, auc_val, auc_test = infer(node_feats, node_types, adjs_pt, pos_val, neg_val, pos_test, neg_test, model_s, model_t)
logging.info("Epoch {}; Train err {}; Val err {}; Val auc {}".format(epoch + 1, train_loss, val_loss, auc_val))
if best_val is None or auc_val > best_val:
best_val = auc_val
final = auc_test
anchor = epoch + 1
logging.info("Best val auc {} at epoch {}; Test auc {}".format(best_val, anchor, final))
def train(node_feats, node_types, adjs, pos_train, neg_train, model_s, model_t, optimizer):
model_s.train()
model_t.train()
optimizer.zero_grad()
out_s = model_s(node_feats, node_types, adjs, archs[args.dataset]["source"][0], archs[args.dataset]["source"][1])
out_t = model_t(node_feats, node_types, adjs, archs[args.dataset]["target"][0], archs[args.dataset]["target"][1])
loss = - torch.mean(F.logsigmoid(torch.mul(out_s[pos_train[:, 0]], out_t[pos_train[:, 1]]).sum(dim=-1)) + \
F.logsigmoid(- torch.mul(out_s[neg_train[:, 0]], out_t[neg_train[:, 1]]).sum(dim=-1)))
loss.backward()
optimizer.step()
return loss.item()
def infer(node_feats, node_types, adjs, pos_val, neg_val, pos_test, neg_test, model_s, model_t):
model_s.eval()
model_t.eval()
with torch.no_grad():
out_s = model_s(node_feats, node_types, adjs, archs[args.dataset]["source"][0], archs[args.dataset]["source"][1])
out_t = model_t(node_feats, node_types, adjs, archs[args.dataset]["target"][0], archs[args.dataset]["target"][1])
#* validation performance
pos_val_prod = torch.mul(out_s[pos_val[:, 0]], out_t[pos_val[:, 1]]).sum(dim=-1)
neg_val_prod = torch.mul(out_s[neg_val[:, 0]], out_t[neg_val[:, 1]]).sum(dim=-1)
loss = - torch.mean(F.logsigmoid(pos_val_prod) + F.logsigmoid(- neg_val_prod))
y_true_val = np.zeros((pos_val.shape[0] + neg_val.shape[0]), dtype=np.long)
y_true_val[:pos_val.shape[0]] = 1
y_pred_val = np.concatenate((torch.sigmoid(pos_val_prod).cpu().numpy(), torch.sigmoid(neg_val_prod).cpu().numpy()))
auc_val = roc_auc_score(y_true_val, y_pred_val)
#* test performance
pos_test_prod = torch.mul(out_s[pos_test[:, 0]], out_t[pos_test[:, 1]]).sum(dim=-1)
neg_test_prod = torch.mul(out_s[neg_test[:, 0]], out_t[neg_test[:, 1]]).sum(dim=-1)
y_true_test = np.zeros((pos_test.shape[0] + neg_test.shape[0]), dtype=np.long)
y_true_test[:pos_test.shape[0]] = 1
y_pred_test = np.concatenate((torch.sigmoid(pos_test_prod).cpu().numpy(), torch.sigmoid(neg_test_prod).cpu().numpy()))
auc_test = roc_auc_score(y_true_test, y_pred_test)
return loss.item(), auc_val, auc_test
if __name__ == '__main__':
main()