-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_and_eval.py
202 lines (148 loc) · 7.44 KB
/
train_and_eval.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
import dgl
import copy
import torch
import numpy as np
from utils import *
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Training for teacher GNNs
def train(model, data, feats, labels, criterion, optimizer, idx_train):
model.train()
logits = model(data, feats)
out = logits.log_softmax(dim=1)
loss = criterion(out[idx_train], labels[idx_train])
optimizer.zero_grad()
loss.backward()
optimizer.step()
return logits, loss.item()
# Training for student MLPs
def train_mini_batch(model, edge_idx, feats, labels, out_t_all, idx_l, criterion_l, criterion_t, optimizer, param):
model.train()
logits = model(None, feats)
out = logits.log_softmax(dim=1)
loss_l = criterion_l(out[idx_l], labels[idx_l])
loss_t = model.edge_distribution_low(edge_idx, out, out_t_all.log_softmax(dim=1), criterion_t)
loss_s = criterion_t(model.edge_distribution_high(edge_idx, logits, param['tau']), model.edge_distribution_high(edge_idx, out_t_all, param['tau']))
if param['ablation_mode'] == 0:
loss = loss_l * param['lamb'] + (loss_t + loss_s) * (1 - param['lamb'])
elif param['ablation_mode'] == 1:
loss = loss_l
elif param['ablation_mode'] == 2:
loss_t = criterion_t(out, out_t_all.log_softmax(dim=1))
loss = loss_l * param['lamb'] + loss_t * (1 - param['lamb'])
elif param['ablation_mode'] == 3:
loss = loss_l * param['lamb'] + loss_t * (1 - param['lamb'])
elif param['ablation_mode'] == 4:
loss = loss_l * param['lamb'] + loss_s * (1 - param['lamb'])
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss_l.item() * param['lamb'], loss_t.item() * (1-param['lamb']), loss_s.item() * (1-param['lamb'])
# Testing for teacher GNNs
def evaluate(model, data, feats, labels, criterion, evaluator, idx_eval):
model.eval()
with torch.no_grad():
logits = model.forward(data, feats)
out = logits.log_softmax(dim=1)
loss = criterion(out[idx_eval], labels[idx_eval])
acc = evaluator(out[idx_eval], labels[idx_eval])
return logits, loss.item(), acc
# Testing for student MLPs
def evaluate_mini_batch(model, feats, labels, criterion, evaluator):
model.eval()
with torch.no_grad():
logits = model.forward(None, feats)
out = logits.log_softmax(dim=1)
loss = criterion(out, labels)
acc = evaluator(out, labels)
return loss.item(), acc
def train_teacher(param, model, g, feats, labels, indices, criterion, evaluator, optimizer):
if param['exp_setting'] == 'tran':
idx_train, idx_val, idx_test = indices
else:
obs_idx_train, obs_idx_val, obs_idx_test, idx_obs, idx_test_ind = indices
obs_feats = feats[idx_obs]
obs_labels = labels[idx_obs]
obs_g = g.subgraph(idx_obs).to(device)
g = g.to(device)
es = 0
val_best = 0
test_val = 0
test_best = 0
for epoch in range(1, param["max_epoch"] + 1):
if param['exp_setting'] == 'tran':
out, loss = train(model, g, feats, labels, criterion, optimizer, idx_train)
_, train_loss, train_acc = evaluate(model, g, feats, labels, criterion, evaluator, idx_train)
_, _, val_acc = evaluate(model, g, feats, labels, criterion, evaluator, idx_val)
_, _, test_acc = evaluate(model, g, feats, labels, criterion, evaluator, idx_test)
else:
out, loss = train(model, obs_g, obs_feats, obs_labels, criterion, optimizer, obs_idx_train)
_, train_loss, train_acc = evaluate(model, obs_g, obs_feats, obs_labels, criterion, evaluator, obs_idx_train)
_, _, val_acc = evaluate(model, obs_g, obs_feats, obs_labels, criterion, evaluator, obs_idx_val)
_, _, test_acc = evaluate(model, g, feats, labels, criterion, evaluator, idx_test_ind)
if test_acc > test_best:
test_best = test_acc
if val_acc >= val_best:
val_best = val_acc
test_val = test_acc
state = copy.deepcopy(model.state_dict())
es = 0
else:
es += 1
if es == 50:
print("Early stopping!")
break
if epoch % 1 == 0:
print("\033[0;30;46m [{}] CLA: {:.5f} | Train: {:.4f}, Val: {:.4f}, Test: {:.4f} | Val Best: {:.4f}, Test Val: {:.4f}, Test Best: {:.4f}\033[0m".format(
epoch, train_loss, train_acc, val_acc, test_acc, val_best, test_val, test_best))
model.load_state_dict(state)
if param['exp_setting'] == 'tran':
out, _, _ = evaluate(model, g, feats, labels, criterion, evaluator, idx_val)
else:
obs_out, _, _ = evaluate(model, obs_g, obs_feats, obs_labels, criterion, evaluator, obs_idx_val)
out, _, _ = evaluate(model, g, feats, labels, criterion, evaluator, idx_test_ind)
out[idx_obs] = obs_out
return out, test_acc, test_val, test_best
def train_student(param, model, g, feats, labels, out_t_all, indices, criterion_l, criterion_t, evaluator, optimizer):
if param['exp_setting'] == 'tran':
idx_train, idx_val, idx_test = indices
idx_l = idx_train
else:
obs_idx_train, obs_idx_val, obs_idx_test, idx_obs, idx_test_ind = indices
obs_idx_l = obs_idx_train
obs_feats = feats[idx_obs]
obs_labels = labels[idx_obs]
obs_out_t = out_t_all[idx_obs]
obs_edge_idx = extract_indices(g.subgraph(idx_obs))
edge_idx = extract_indices(g)
es = 0
val_best = 0
test_val = 0
test_best = 0
for epoch in range(1, param["max_epoch"] + 1):
if param['exp_setting'] == 'tran':
loss_l, loss_t, loss_s = train_mini_batch(model, edge_idx, feats, labels, out_t_all, idx_l, criterion_l, criterion_t, optimizer, param)
loss = loss_l + loss_t + loss_s
train_loss, train_acc = evaluate_mini_batch(model, feats[idx_train], labels[idx_train], criterion_l, evaluator)
_, val_acc = evaluate_mini_batch(model, feats[idx_val], labels[idx_val], criterion_l, evaluator)
_, test_acc = evaluate_mini_batch(model, feats[idx_test], labels[idx_test], criterion_l, evaluator)
else:
loss_l, loss_t, loss_s = train_mini_batch(model, obs_edge_idx, obs_feats, obs_labels, obs_out_t, obs_idx_l, criterion_l, criterion_t, optimizer, param)
loss = loss_l + loss_t + loss_s
train_loss, train_acc = evaluate_mini_batch(model, obs_feats[obs_idx_train], obs_labels[obs_idx_train], criterion_l, evaluator)
_, val_acc = evaluate_mini_batch(model, obs_feats[obs_idx_val], obs_labels[obs_idx_val], criterion_l, evaluator)
_, test_acc = evaluate_mini_batch(model, feats[idx_test_ind], labels[idx_test_ind], criterion_l, evaluator)
if test_acc > test_best:
test_best = test_acc
if val_acc >= val_best:
val_best = val_acc
test_val = test_acc
es = 0
else:
es += 1
if es == 50:
print("Early stopping!")
break
if epoch % 1 == 0:
print("\033[0;30;46m [{}] CLA: {:.5f}, KD: {:.5f}, SK: {:.5f}, Total: {:.5f} | Train: {:.4f}, Val: {:.4f}, Test: {:.4f} | Val Best: {:.4f}, Test Val: {:.4f}, Test Best: {:.4f}\033[0m".format(
epoch, loss_l, loss_t, loss_s, loss, train_acc, val_acc, test_acc, val_best, test_val, test_best))
return test_acc, test_val, test_best