-
Notifications
You must be signed in to change notification settings - Fork 1
/
maml.py
128 lines (99 loc) · 3.84 KB
/
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
import torch
from torch import optim
import joblib
import os
import argparse
import numpy as np
from model import CNNText
from util import PyTorchParameterList2NPArrayList
from loss import EDL_Loss
# config #
inner_rate = 0.01
outer_rate = 0.1
n_subsets = 5
ntrain = 10
loss_fn = EDL_Loss()
print('[Start maml ...]')
parser = argparse.ArgumentParser()
parser.add_argument("--file_path", help="saving root path of raw data", default='./test')
parser.add_argument("--seed", help="reproducible experiment with seeds", type=int)
parser.add_argument('--is_doc2vec', help='use doc2vec clustering', default='0')
parser.add_argument("--out_dim", help="output dimension", type=int, default=6)
parser.add_argument('--niterations', help='number of iterations', default=1000, type=int)
args = parser.parse_args()
RandomGenerator = np.random.RandomState(args.seed)
if args.is_doc2vec is '1':
data_file_path = os.path.join(args.file_path, 'data/data_doc2vec.pkl')
elif args.is_doc2vec is '0':
data_file_path = os.path.join(args.file_path, 'data/data.pkl')
else:
raise ValueError
[train_tasks, test_tasks, vocabulary, pretrained_embeddings, X_test, y_test] = joblib.load(data_file_path)
vocab_size = len(vocabulary)
sentence_len = X_test.shape[1]
model = CNNText(vocab_size, sentence_len, pretrained_embeddings, args.out_dim).cuda()
meta_optim = optim.SGD(model.parameters(), lr=outer_rate)
num_train_tasks = len(train_tasks)
for iteration in range(args.niterations + 1):
total_loss = None
for i in range(n_subsets):
model.train()
k = RandomGenerator.randint(0, high=num_train_tasks)
task = train_tasks[k]
Xtrain, ytrain = task.get_train()
Xtest, ytest = task.get_test()
Xtrain = torch.from_numpy(Xtrain).long().cuda()
ytrain = torch.from_numpy(ytrain).float().cuda()
Xtest = torch.from_numpy(Xtest).long().cuda()
ytest = torch.from_numpy(ytest).float().cuda()
m = len(Xtrain)
fast_weights = None
inds = RandomGenerator.permutation(m)
mbinds = inds[0:ntrain]
ypred, _ = model(Xtrain[mbinds], vars=fast_weights)
ypred = ypred.cuda()
# choose = RandomGenerator.randint(0, high=5)
# if choose == 0:
# loss_fn = KL_Loss()
# elif choose == 1:
# loss_fn = CE_Loss()
# elif choose == 2:
# loss_fn = EDL_Loss()
# elif choose == 3:
# loss_fn = Euclidean_Loss()
# elif choose == 4:
# loss_fn = Cosine_Loss()
loss = loss_fn(ytrain[mbinds], ypred)
grad = torch.autograd.grad(loss, model.parameters())
fast_weights = list(map(lambda p: p[1] - inner_rate * p[0], zip(grad, model.parameters())))
n = len(Xtest)
inds = RandomGenerator.permutation(n)
model.eval()
ypred, _ = model(Xtest[inds], vars=fast_weights)
ypred = ypred.cuda()
y = ytest[inds]
if total_loss is None:
total_loss = loss_fn(y, ypred)
else:
tmp_loss = loss_fn(y, ypred)
total_loss += tmp_loss
total_loss /= n_subsets
model.train()
meta_optim.zero_grad()
total_loss.backward()
meta_optim.step()
print('[{:d}/{:d}] ... ...'.format(iteration, args.niterations))
if iteration % 1000 == 0:
param_models = []
param = PyTorchParameterList2NPArrayList(model.parameters())
param_models.append(param)
if args.is_doc2vec is '1':
directory = os.path.join(args.file_path, 'trained_doc2vec')
elif args.is_doc2vec is '0':
directory = os.path.join(args.file_path, 'trained')
else:
raise ValueError
if not os.path.exists(directory):
os.makedirs(directory)
joblib.dump([param_models], os.path.join(directory, 'store-' + str(iteration) + '.pkl'))
print('[Finish maml ...]')