/
train_sgf.py
203 lines (180 loc) · 7.13 KB
/
train_sgf.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
import time
import random
import argparse
import numpy as np
import pickle as pkl
import torch
import torch.nn.functional as F
import torch.optim as optim
from model import SGF
import uuid
from collections import Counter
from sklearn.metrics import confusion_matrix
from rayleigh import rayleigh_quotient, rayleigh_sub
from utils import load_data, accuracy, get_coeff
# Training settings
parser = argparse.ArgumentParser()
parser.add_argument('--data', default='cora', help='Dataset')
parser.add_argument('--seed', type=int, default=0, help='Random seed')
parser.add_argument('--epochs', type=int, default=1500, help='Max epoch.')
parser.add_argument('--lr', type=float, default=0.01, help='Learning rate')
parser.add_argument('--wd', type=float, default=0.0005, help='Weight decay')
parser.add_argument('--layer', type=int, default=16, help='Number of layers')
parser.add_argument('--hidden', type=int, default=64, help='Hidden dimension')
parser.add_argument('--dropout', type=float, default=0.7, help='Dropout rate')
parser.add_argument('--patience', type=int, default=200, help='Patience')
parser.add_argument('--cuda', action="store_true", default=False, help='Train on CPU or GPU')
parser.add_argument('--gpu_id', type=int, default=0, help='GPU device id')
parser.add_argument('--test', action='store_true', default=False, help='evaluation on test set.')
parser.add_argument('--test_study', action='store_true', default=False, help='print info on the test result.')
parser.add_argument("--log_period", type=int, default=50, help="Log every x epochs")
parser.add_argument("--split", type=str, default="0.6_0.2_0.2")
parser.add_argument("--use_laplacian", action="store_true", default=False)
parser.add_argument("--analyze_filter", action="store_true", default=False)
parser.add_argument("--perturbate_edges", type=float, default=0.0)
args = parser.parse_args()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
# Sklearn can't split this one so we have to use the
# default splits
if args.data in ["cornell", "texas"]:
args.split = "0.6"
# Load data
if args.perturbate_edges > 0.0:
print("INFO: Perturbate edges")
data_package = load_data(args.data,
["AugNormAdj", "SymNormLap"],
split=args.split, rs=args.seed,
pf=args.perturbate_edges)
else:
data_package = load_data(args.data,
["AugNormAdj", "SymNormLap"],
split=args.split, rs=args.seed)
_, normed_adjs, features, labels, idx_train, idx_val, idx_test = data_package
adj, L = normed_adjs
if args.use_laplacian:
adj = L
### This is only used to print rayleigh loss, not for training!
### The training process strictly use only idx_train
est_rayleigh = rayleigh_sub(labels, L,
torch.cat((idx_train, idx_val)),
one_hot=True)
######
# Setup device
if args.cuda:
cudaid = "cuda:"+str(args.gpu_id)
else:
cudaid = "cpu"
device = torch.device(cudaid)
features = features.to(device)
adj = adj.to(device)
L = L.to(device)
checkpt_file = 'checkpoints/'+uuid.uuid4().hex[:4]+'-'+args.data+'.pt'
poly_file = 'checkpoints/'+uuid.uuid4().hex[:4]+'-'+args.data+'.poly'
print(cudaid, checkpt_file)
if args.test_study:
print(poly_file)
model = SGF(nfeat=features.shape[1],
nlayers=args.layer,
nhidden=args.hidden,
nclass=int(labels.max()) + 1,
dropout=args.dropout).to(device)
opt_config = [
{'params': model.params1, 'weight_decay': args.wd, 'lr': args.lr},
{'params': model.params2, 'weight_decay': args.wd, 'lr': args.lr/4},
]
optimizer = optim.Adam(opt_config)
def train():
model.train()
optimizer.zero_grad()
output = model(features, adj)
acc_train = accuracy(output[idx_train], labels[idx_train].to(device))
loss_train = F.nll_loss(output[idx_train], labels[idx_train].to(device))
loss_train.backward()
optimizer.step()
return loss_train.item(), acc_train.item()
def validate():
model.eval()
with torch.no_grad():
output = model(features, adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val].to(device))
acc_val = accuracy(output[idx_val], labels[idx_val].to(device))
return loss_val.item(), acc_val.item()
def test():
model.load_state_dict(torch.load(checkpt_file))
model.eval()
with torch.no_grad():
output = model(features, adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test].to(device))
acc_test = accuracy(output[idx_test], labels[idx_test].to(device))
return loss_test.item(), acc_test.item()
def filt_eval():
model.eval()
with torch.no_grad():
out = model(features, adj)
out_r = rayleigh_quotient(out, L)
loss_r = F.l1_loss(out_r, est_rayleigh.to(device))
return loss_r.item()
def test_study():
model.load_state_dict(torch.load(checkpt_file))
model.eval()
with torch.no_grad():
output = model(features, adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test].to(device))
preds = output[idx_test].max(1)[1].type_as(labels[idx_test])
correct = preds.eq(labels[idx_test]).double()
correct_count = correct.sum()
acc_test = correct_count / len(idx_test)
wrong_labels = (1+labels[idx_test]) * (1-correct)
wrong_labels = Counter(wrong_labels.cpu().numpy().astype(int))
test_labels = Counter(labels[idx_test].cpu().numpy().astype(int))
for k, v in test_labels.items():
print(k, 1 - wrong_labels[k+1]/v)
alphas = [f.alpha1.item() for f in model.filters]
betas = [f.alpha2.item() for f in model.filters]
poly_coeffs = get_coeff(alphas, betas)
print(poly_coeffs)
with open(poly_file, "wb") as f:
pkl.dump(poly_coeffs, f)
print(confusion_matrix(labels[idx_test], preds))
return loss_test.item(), acc_test.item()
t_total = time.time()
c = 0
best = 0
best_epoch = 0
acc = 0
for epoch in range(args.epochs):
# _ = filt_train()
loss_tra, acc_tra = train()
loss_val, acc_val = validate()
loss_filt = filt_eval()
if (epoch+1)%args.log_period == 0 or epoch == 0:
print('Epoch:{:04d}'.format(epoch+1),
'loss:{:.3f}'.format(loss_tra),
'acc:{:.2f}'.format(acc_tra*100),
' | ',
'loss:{:.3f}'.format(loss_val),
'acc:{:.2f}'.format(acc_val*100),
' | ',
'f_loss:{:.3f}'.format(loss_filt))
if args.analyze_filter and epoch % 20 == 0:
torch.save(model.state_dict(), "{}_{}.pt".format(checkpt_file[:-3], epoch))
if acc_val > best:
best = acc_val
best_epoch = epoch
acc = acc_val
torch.save(model.state_dict(), checkpt_file)
c = 0
else:
c += 1
if c == args.patience:
break
if args.test:
acc = test()[1]
if args.test_study:
acc = test_study()[1]
print("Train cost: {:.4f}s".format(time.time() - t_total))
print('Load {}th epoch'.format(best_epoch))
print("Test" if args.test or args.test_study else "Val","acc.:{:.1f}".format(acc*100))