-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_ICL_MAML.py
157 lines (133 loc) · 6.55 KB
/
train_ICL_MAML.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 torch
import torch_geometric.data
import os
import numpy as np
from criterion import loss_cal_ICL_old, loss_cal_ICL_new
from augmentation import aug_graph, aug_img
import time
from util import evaluate_embedding
from collections import OrderedDict
def train_ICL_MAML(data, old_data, new_data, encoder, args):
# optimizer
optimizer = torch.optim.Adam(encoder.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# data loader
if args.mode == 'graph':
loader = torch_geometric.data.DataLoader(data, batch_size=args.batch_size, shuffle=True)
old_loader = torch_geometric.data.DataLoader(old_data, batch_size=args.batch_size, shuffle=True)
new_loader = torch_geometric.data.DataLoader(new_data, batch_size=args.batch_size, shuffle=True)
aug = aug_graph
elif args.mode == 'cv':
loader = torch.utils.data.DataLoader(data, batch_size=args.batch_size, shuffle=True)
old_loader = torch.utils.data.DataLoader(old_data, batch_size=args.batch_size, shuffle=True)
new_loader = torch.utils.data.DataLoader(new_data, batch_size=args.batch_size, shuffle=True)
aug = aug_img
# train data
final_epoch = 0
results = []
total_time = 0
count = 0
pkl_save = '{}.{}.{}.pkl'.format(args.dataset, args.alpha, time.time())
for epoch in range(args.epochs):
loss_all = 0
# train
encoder.train()
start_time = time.time()
old_step = max(int(len(old_loader) / len(new_loader)), 1)
for new_batch in new_loader:
fast_weights = OrderedDict(encoder.named_parameters())
# meta-train
for step in range(old_step):
try:
old_batch = old_iter.next()
except:
old_iter = iter(old_loader)
old_batch = old_iter.next()
old_batch, old_batch_aug = aug(old_batch, np.random.randint(3))
optimizer.zero_grad()
old_out = encoder(old_batch, fast_weights)
old_out_aug = encoder(old_batch_aug, fast_weights)
try:
if args.mode == 'graph':
new4old_batch = new4old_iter.next()
elif args.mode == 'cv':
new4old_batch = new4old_iter.next()[0][0]
except:
if args.mode == 'graph':
new4old_iter = iter(torch_geometric.data.DataLoader(new_data, batch_size=args.batch_size-1, shuffle=True))
new4old_batch = new4old_iter.next()
elif args.mode == 'cv':
new4old_iter = iter(torch.utils.data.DataLoader(new_data, batch_size=args.batch_size-1, shuffle=True))
new4old_batch = new4old_iter.next()[0][0]
new4old_out = encoder(new4old_batch, fast_weights)
loss = loss_cal_ICL_old(old_out, old_out_aug, new4old_out, args.alpha)
gradients = torch.autograd.grad(loss, fast_weights.values())
# update weights manually
fast_weights = OrderedDict(
(name, param - args.lr * grad)
for ((name, param), grad) in zip(fast_weights.items(), gradients)
)
# meta-test
if args.mode == 'graph':
batch_size = new_batch.num_graphs
elif args.mode == 'cv':
batch_size = len(new_batch[0][0])
old_num = int((1 - args.alpha) * (batch_size-1))
new_num = batch_size-1 - old_num
new_batch, new_batch_aug = aug(new_batch, np.random.randint(3))
optimizer.zero_grad()
new_out = encoder(new_batch, fast_weights)
new_out_aug = encoder(new_batch_aug, fast_weights)
if old_num == 0:
old_out = None
else:
try:
if args.mode == 'graph':
old_batch = old4new_iter.next()
elif args.mode == 'cv':
old_batch = old4new_iter.next()[0][0]
except:
if args.mode == 'graph':
old4new_iter = iter(torch_geometric.data.DataLoader(old_data, batch_size=old_num, shuffle=True))
old_batch = old4new_iter.next()
elif args.mode == 'cv':
old4new_iter = iter(torch.utils.data.DataLoader(old_data, batch_size=old_num, shuffle=True))
old_batch = old4new_iter.next()[0][0]
old_out = encoder(old_batch, fast_weights)
loss = loss_cal_ICL_new(new_out, new_out_aug, old_out, new_num)
loss_all += loss.item() #* data.num_graphs
loss.backward()
optimizer.step()
if torch.cuda.is_available():
torch.cuda.empty_cache()
end_time = time.time()
# get result
average_loss = loss_all / (len(new_loader))
epoch_time = end_time - start_time
total_time += epoch_time
result = [epoch, average_loss, epoch_time, total_time]
results.append(result)
if epoch % 10 == 0:
print('Epoch {:<5}, Loss {:.4f}, time {:.4f}, total time {:.4f}'.format(*result))
if results[final_epoch][1] >= average_loss:
final_epoch = epoch
try:
torch.save(encoder.module.state_dict(), pkl_save)
except:
torch.save(encoder.state_dict(), pkl_save)
count = 0
else:
count += 1
if count == args.patience:
break
print('Eval!')
encoder.load_state_dict(torch.load(pkl_save))
encoder.eval()
embs, y = encoder.get_embeddings(loader)
acc = evaluate_embedding(embs, y)
embs, y = encoder.get_embeddings(old_loader)
acc_old = evaluate_embedding(embs, y)
embs, y = encoder.get_embeddings(new_loader)
acc_new = evaluate_embedding(embs, y)
acc_new = evaluate_embedding(embs, y)
print('Final Epoch {:<5}, Loss {:.4f}, time {:.4f}, total time {:.4f}, acc {:.4f}, acc-old {:.4f}, acc-new {:.4f}'.format(*(results[final_epoch]+[acc, acc_old, acc_new])))
os.remove(pkl_save)