-
Notifications
You must be signed in to change notification settings - Fork 0
/
alds.py
225 lines (193 loc) · 7.53 KB
/
alds.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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
import torch
import torch.nn as nn
import argparse
from model import ALDS_model
from loss_ALDS import ALDSloss
from dataset import TrainData
import torch.nn.functional as F
from torch.utils.data import DataLoader
import os
import numpy as np
import pandas as pd
import random
from sklearn.utils import shuffle
import cfg
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline
from sklearn.metrics import roc_auc_score as AUC
from sklearn.metrics import roc_curve
from sklearn.preprocessing import label_binarize
import time
import matplotlib.pyplot as plt
import scipy.io as sio
from sklearn import preprocessing
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
parser = argparse.ArgumentParser(description='Hyper-parameters')
parser.add_argument('-lr', default=1e-3, type=float)
parser.add_argument('-pretrain_epochs', default=100, type=int, help='Set epochs for training')
args = parser.parse_args()
epochs = args.pretrain_epochs
learning_rate = args.lr
_C = cfg._C_musk
# Wavaform Dataset
dataset_dir = _C['dataset_dir']
dataset_size = _C['train_size']
sketch_size = _C['sketch_size']
batch_size = _C['batch_size']
print('reading...')
dataframe = np.load(dataset_dir)
print('readok')
# Save
save_dir = _C['save_model_dir']
# GPU
device_ids = [Id for Id in range(torch.cuda.device_count())]
device_ids = [0, 1, 2, 3]
def generate_seed(seed):
torch.cuda.manual_seed(cfg.seed)
random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
def train(alpha, beta, gamma, lamb):
def norm_21(x):
x = x * x
x = torch.sum(x, 0)
x = torch.sqrt(x)
return torch.sum(x)
def norm_inf(x):
x = torch.abs(x)
x = torch.max(x, 0)[0]
print(x.size())
return torch.sum(x)
ALDS = ALDS_model(_C['feature_in'], origin_size=dataset_size, sketch_size=sketch_size, _C=_C, fine_tune=False)
ALDS = ALDS.cuda()
train_data_loader = DataLoader(TrainData(dataset_dir, _C['SNR']), batch_size=_C['batch_size'], shuffle=False, num_workers=0)
# Define optimizer
opt = torch.optim.Adam(ALDS.parameters(), lr=learning_rate)
#define losses #
CELoss = nn.CrossEntropyLoss()
MSELoss = nn.MSELoss(reduction='sum')
print("Total number of paramerters in networks is {} ".format(sum(x.numel() for x in ALDS.parameters())))
# pretrain
for epoch in range(epochs):
for batch_id, train_data in enumerate(train_data_loader):
# Get train_data
X, X_label, Fake, Fake_label = train_data
X, X_label, Fake, Fake_label = X.cuda(), X_label.cuda(), Fake.cuda(), Fake_label.cuda()
opt.zero_grad()
X_prime, score_x, score_fake = ALDS(X, Fake)
loss_r = MSELoss(X_prime, X)
loss_c_true = CELoss(score_x, X_label)
loss_c_false = CELoss(score_fake, Fake_label)
loss_c = loss_c_true + loss_c_false
total_loss = loss_r + alpha*loss_c
total_loss.backward()
opt.step()
if epoch % 2 == 0:
print("epoch:{}, loss_r:{}, loss_c:{}, total_loss:{}"\
.format(epoch, loss_r.item(), loss_c.item(), total_loss.item()))
state_dict = ALDS.state_dict()
torch.save(state_dict, os.path.join(save_dir, 'pretrain'))
print('---------------------- fine-tuning -----------------------')
#fine-tuning
ALDS = ALDS_model(_C['feature_in'], origin_size=dataset_size, sketch_size=sketch_size, fine_tune=True, _C=_C)
ALDS = ALDS.cuda()
state_dict = torch.load(os.path.join(save_dir, 'pretrain'))
ALDS.load_state_dict(state_dict)
opt = torch.optim.Adam(ALDS.parameters(), lr=learning_rate / 2)
# use whole dataset as a batch
batch_size = dataset_size
print("batch_size:{}".format(batch_size))
train_data_loader = DataLoader(TrainData(dataset_dir, _C['SNR']), \
batch_size=batch_size, shuffle=False, num_workers=0
)
for epoch in range(20):
for batch_id, train_data in enumerate(train_data_loader):
X, X_label, Fake, Fake_label = train_data
X, X_label, Fake, Fake_label = X.cuda(), X_label.cuda(), Fake.cuda(), Fake_label.cuda()
opt.zero_grad()
S, S_prime, Z, Z_prime, X_prime, score_x, score_fake, Q = ALDS(X, Fake)
loss_c_true = CELoss(score_x, X_label)
loss_c_false = CELoss(score_fake, Fake_label)
loss_c = loss_c_true + loss_c_false
Loss = ALDSloss()
loss = Loss(S, S_prime, Z, Z_prime, X, X_prime, Q, lamb)
loss_S = loss['loss_S']
loss_recS = loss['loss_recS']
norm_q = loss['norm_q']
loss_Z = loss['loss_Z']
loss_r = loss['loss_r']
total_loss = beta*loss_S + loss_r + gamma*loss_Z + alpha*loss_c
total_loss.backward()
opt.step()
print("epoch:{}, loss_recS:{}, norm:{}, loss_r:{}, loss_c:{}, total_loss:{}"\
.format(epoch, loss_recS.item(), norm_q.item(), loss_r.item(), loss_c.item(), total_loss.item()))
test(ALDS.state_dict(), state_dict)
def test(ft_state_dict, pre_state_dict):
query_list = np.linspace(50, 600, 12, dtype=np.int32)
# query_list = np.linspace(25, 300, 12, dtype=np.int32)
ALDS = ALDS_model(_C['feature_in'], origin_size=dataset_size, sketch_size=sketch_size, fine_tune=True, _C=_C, test=True)
ALDS = ALDS.cuda()
ALDS.load_state_dict(pre_state_dict)
def select(Q, query):
dataset = np.array(dataframe, dtype=np.float32)[0:_C['train_size']]
_Q = Q.cpu().numpy()
_Q = np.sum(_Q, 0)
idx = np.argsort(-_Q)
np.save(os.path.join(save_dir, 'samples_index.npy'), idx)
# sio.savemat(os.path.join(save_dir,'order{}.mat'.format(seq)),{'order':idx})
idx = idx[0:query]
n, d = dataset.shape
datas = dataset[idx, :d - 1]
label = dataset[idx, -1]
label = label.reshape(label.shape[0], 1)
datas = ALDS(torch.from_numpy(datas).cuda())
datas = datas.cpu().detach().numpy()
datas = np.concatenate((datas, label), axis=1)
return datas
def calc_acc(clf, X, y):
y_test = clf.predict(X)
acc = sum(y_test == y) / len(y)
return acc
def calc_auc(clf, X, y):
n_class = np.unique(y).shape[0]
print(n_class)
y_one_hot = label_binarize(y, np.arange(n_class))
print(y_one_hot)
y_score = clf.predict_proba(X)
auc = AUC(y_one_hot, y_score, average='micro')
return auc
def svm_train(query):
dataset = select(ft_state_dict['Q'], query)
n, d = dataset.shape
X = dataset[:, :d - 1]
y = dataset[:, -1]
from sklearn.svm import LinearSVC, SVC
clf = SVC(C=100, kernel='linear')
start = time.time()
clf.fit(X, y)
end = time.time()
return clf
def svm_test(clf):
dataset = np.array(dataframe, dtype=np.float32)[_C['test_size']:,:]
n, d = dataset.shape
X = dataset[:, :d - 1]
y = dataset[:, -1]
X = ALDS(torch.from_numpy(X).cuda())
X = X.cpu().detach().numpy()
acc = calc_acc(clf, X, y)
return acc
acc = []
for query in query_list:
start = time.time()
clf = svm_train(query)
train_time = time.time()
acc.append(svm_test(clf))
print(acc)
def main():
# hyper-parameters
alpha = 0.1
beta = 1
gamma = 0.01
lamb = 10
train(alpha, beta, gamma, lamb)
if __name__ == '__main__':
main()