-
Notifications
You must be signed in to change notification settings - Fork 0
/
NASH.py
204 lines (167 loc) · 7.03 KB
/
NASH.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
import torch.autograd as autograd
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from tqdm import tqdm
from utils.utils import *
from utils.datasets import *
def get_config():
config = {
"device": torch.device("cuda:0"),
"dataset": "ng20.tfidf",
"bit": 8,
"dropout": 0.1,
"batch_size": 64,
"epoch": 300,
"optimizer": {"type": optim.Adam, "optim_params": {"lr": 0.001}},
"stop_iter": 10
}
return config
class LBSign(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
return torch.sign(input)
@staticmethod
def backward(ctx, grad_output):
return grad_output
class NASH(nn.Module):
def __init__(self, vocabSize, latentDim, device, dropoutProb=0.) :
super(NASH, self).__init__()
# self.hidden_dim = 1000
# according to paper, we set the hidden_dim as 500
self.hidden_dim = 500
self.vocabSize = vocabSize
self.latentDim = latentDim
self.dropoutProb = dropoutProb
self.device = device
# encoder network
self.encoder = nn.Sequential(
nn.Linear(self.vocabSize, self.hidden_dim), nn.ReLU(inplace=True),
nn.Linear(self.hidden_dim, self.hidden_dim), nn.ReLU(inplace=True),
nn.Dropout(p=dropoutProb))
self.h_to_mu = nn.Sequential(
nn.Linear(self.hidden_dim, self.latentDim), nn.Sigmoid())
# decoder network
self.decoder = nn.Sequential(nn.Linear(self.latentDim, self.vocabSize),
nn.LogSoftmax(dim=1))
# noise network
self.sigma = nn.Sequential(
nn.Linear(self.latentDim, self.latentDim)
, nn.Sigmoid()
, nn.Dropout(self.dropoutProb)
)
# self.deterministic = config.deterministic
# self.mu = None
# if not self.deterministic :
# self.mu = torch.rand(config.output_size)
# if config.use_cuda :
# self.mu = self.mu.cuda()
def binarization(self, mu, isStochastic):
lb_sign = LBSign.apply
if isStochastic:
thresh = torch.FloatTensor(mu.size()).uniform_().to(self.device)
return (lb_sign(mu - thresh) + 1) / 2
else:
return (lb_sign(mu - 0.5) + 1) / 2
def encode(self, doc_mat, isStochastic):
h = self.encoder(doc_mat)
mu = self.h_to_mu(h)
z = self.binarization(mu, isStochastic)
return z, mu
def forward(self, document_mat, isStochastic, integration='sum'):
z, mu = self.encode(document_mat, isStochastic)
# add noise from mu
z_noise = z + self.sigma(mu)
prob_w = self.decoder(z_noise)
prob_w = self.decoder(z)
return prob_w, mu
def get_name(self):
return "NASH"
@staticmethod
def calculate_KL_loss(mu):
thresh = 1e-20 * torch.ones(mu.size()).cuda()
KLD_element = mu * torch.log(torch.max(mu * 2, thresh)) + (
1 - mu) * torch.log(torch.max((1 - mu) * 2, thresh))
KLD = torch.sum(KLD_element, dim=1)
KLD = torch.mean(KLD)
return KLD
@staticmethod
def compute_reconstr_loss(logprob_word, doc_mat):
return -torch.mean(torch.sum(logprob_word * doc_mat, dim=1))
def get_binary_code(self, train, test, isStochastic):
train_zy = [(self.encode(xb.to(self.device), isStochastic)[0], yb)
for xb, yb in train]
train_z, train_y = zip(*train_zy)
train_z = torch.cat(train_z, dim=0)
train_y = torch.cat(train_y, dim=0)
test_zy = [(self.encode(xb.to(self.device), isStochastic)[0], yb)
for xb, yb in test]
test_z, test_y = zip(*test_zy)
test_z = torch.cat(test_z, dim=0)
test_y = torch.cat(test_y, dim=0)
train_b = train_z.type(torch.cuda.ByteTensor)
test_b = test_z.type(torch.cuda.ByteTensor)
del train_z
del test_z
return train_b, test_b, train_y, test_y
def train_val(config):
bit = config["bit"]
device = config["device"]
dataset, data_fmt = config["dataset"].split('.')
batch_size = config["batch_size"]
if dataset in ['reuters', 'tmc', 'rcv1']:
single_label_flag = False
else:
single_label_flag = True
if single_label_flag:
train_set = SingleLabelTextDataset('datasets/{}'.format(dataset), subset='train', bow_format=data_fmt, download=True)
test_set = SingleLabelTextDataset('datasets/{}'.format(dataset), subset='test', bow_format=data_fmt, download=True)
else:
train_set = MultiLabelTextDataset('datasets/{}'.format(dataset), subset='train', bow_format=data_fmt, download=True)
test_set = MultiLabelTextDataset('datasets/{}'.format(dataset), subset='test', bow_format=data_fmt, download=True)
train_loader = torch.utils.data.DataLoader(dataset=train_set, batch_size=batch_size, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_set, batch_size=batch_size, shuffle=True)
num_bits = bit
num_features = train_set[0][0].size(0)
model = NASH(num_features, num_bits, dropoutProb=0.1, device=device)
model.to(device)
num_epochs = config["epoch"]
optimizer = config["optimizer"]["type"](model.parameters(), lr=config["optimizer"]["optim_params"]["lr"])
kl_weight = 0.
kl_step = 5e-6
best_precision = 0
best_precision_epoch = 0
step_count = 0
for epoch in tqdm(range(config["epoch"])):
avg_loss = []
for step, (xb, yb) in enumerate(train_loader):
xb = xb.to(device)
yb = yb.to(device)
logprob_w, mu = model(xb, isStochastic=True)
kl_loss = NASH.calculate_KL_loss(mu)
reconstr_loss = NASH.compute_reconstr_loss(logprob_w, xb)
loss = reconstr_loss + kl_weight * kl_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
kl_weight = min(kl_weight + kl_step, 1.)
avg_loss.append(loss.item())
with torch.no_grad():
train_b, test_b, train_y, test_y = model.get_binary_code(train_loader, test_loader, isStochastic=True)
retrieved_indices = retrieve_topk(test_b.to(device), train_b.to(device), topK=100)
prec = compute_precision_at_k(retrieved_indices, test_y.to(device), train_y.to(device), topK=100, is_single_label=single_label_flag)
print("precision at 100: {:.4f}".format(prec))
if prec.item() > best_precision:
best_precision = prec.item()
best_precision_epoch = epoch + 1
step_count = 0
else:
step_count += 1
if step_count >= config["stop_iter"]:
break
tqdm.write(f'Epoch {epoch+1}/{config["epoch"]} - Current Precision: {prec.item():.4f} - Best Precision: {best_precision:.4f} [{best_precision_epoch}]')
if __name__ == "__main__":
config = get_config()
train_val(config)