forked from kimhc6028/soft-decision-tree
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
267 lines (231 loc) · 10.1 KB
/
model.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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
import os
import pickle
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.optim.lr_scheduler as lrs
from torch.autograd import Variable
from sklearn.metrics import roc_auc_score
class InnerNode:
def __init__(self, depth, args):
self.args = args
self.fc = nn.Linear(self.args.input_dim, 1)
# beta = torch.randn(1)
# if self.args.cuda:
# beta = beta.cuda()
# self.beta = nn.Parameter(beta)
self.leaf = False
self.prob = None
self.leaf_accumulator = []
self.lmbda = self.args.lmbda * 2 ** (-depth)
self.build_child(depth)
self.penalties = []
def reset(self):
self.leaf_accumulator = []
self.penalties = []
self.left.reset()
self.right.reset()
def build_child(self, depth):
if depth < self.args.max_depth:
self.left = InnerNode(depth + 1, self.args)
self.right = InnerNode(depth + 1, self.args)
else:
self.left = LeafNode(self.args)
self.right = LeafNode(self.args)
def forward(self, x):
# return F.sigmoid(self.beta * self.fc(x))
return F.sigmoid(self.fc(x))
# def select_next(self, x):
# prob = self.forward(x)
# if prob < 0.5:
# return self.left, prob
# else:
# return self.right, prob
def cal_prob(self, x, path_prob):
self.prob = self.forward(x) # probability of selecting right node
self.path_prob = path_prob
left_leaf_accumulator = self.left.cal_prob(x, path_prob * (1 - self.prob))
right_leaf_accumulator = self.right.cal_prob(x, path_prob * self.prob)
self.leaf_accumulator.extend(left_leaf_accumulator)
self.leaf_accumulator.extend(right_leaf_accumulator)
return self.leaf_accumulator
def get_penalty(self):
penalty = (torch.sum(self.prob * self.path_prob) / torch.sum(self.path_prob), self.lmbda)
if not self.left.leaf:
left_penalty = self.left.get_penalty()
right_penalty = self.right.get_penalty()
self.penalties.append(penalty)
self.penalties.extend(left_penalty)
self.penalties.extend(right_penalty)
return self.penalties
class LeafNode:
def __init__(self, args):
self.args = args
self.param = torch.randn(self.args.output_dim)
if self.args.cuda:
self.param = self.param.cuda()
self.param = nn.Parameter(self.param)
self.leaf = True
self.softmax = nn.Softmax()
def forward(self):
return self.softmax(self.param.view(1, -1))
def reset(self):
pass
def cal_prob(self, x, path_prob):
Q = self.forward()
Q = Q.expand((path_prob.size()[0], self.args.output_dim))
return [[path_prob, Q]]
class SoftDecisionTree(nn.Module):
def __init__(self, args):
super(SoftDecisionTree, self).__init__()
self.args = args
self.root = InnerNode(1, self.args)
self.collect_parameters() # collect parameters and modules under root node
self.optimizer = optim.SGD(self.parameters(), lr=self.args.lr, momentum=self.args.momentum)
# self.optimizer = optim.Adam(self.parameters(), lr=self.args.lr)
self.scheduler = lrs.ReduceLROnPlateau(self.optimizer, factor=0.9, patience=50, verbose=True)
self.test_acc = []
self.define_extras(self.args.batch_size)
self.best_auc_roc = 0.
self.alpha = torch.FloatTensor([0.001])
self.alpha = nn.Parameter(self.alpha, requires_grad=False)
self.bn = nn.BatchNorm1d(self.args.input_dim)
def define_extras(self, batch_size):
# define target_onehot and path_prob_init batch size, because these
# need to be defined according to batch size, which can be differ
self.target_onehot = torch.FloatTensor(batch_size, self.args.output_dim)
self.target_onehot = Variable(self.target_onehot)
self.path_prob_init = Variable(torch.ones(batch_size, 1))
if self.args.cuda:
self.target_onehot = self.target_onehot.cuda()
self.path_prob_init = self.path_prob_init.cuda()
def cal_loss(self, x, y):
batch_size = y.size()[0]
leaf_accumulator = self.root.cal_prob(x, self.path_prob_init)
loss = 0.
max_prob = [-1. for _ in range(batch_size)]
max_Q = [torch.zeros(self.args.output_dim) for _ in range(batch_size)]
for (path_prob, Q) in leaf_accumulator:
TQ = torch.bmm(y.view(batch_size, 1, self.args.output_dim),
torch.log(Q).view(batch_size, self.args.output_dim, 1)).view(-1, 1)
loss += path_prob * TQ
path_prob_numpy = path_prob.cpu().data.numpy().reshape(-1)
for i in range(batch_size):
if max_prob[i] < path_prob_numpy[i]:
max_prob[i] = path_prob_numpy[i]
max_Q[i] = Q[i]
loss = loss.mean()
loss = -loss
# Let's add L1 regularization
l1_reg = 0
for layer in self.module_list:
layer_params = list(layer.parameters())
l1_reg += layer_params[0].abs().sum() + layer_params[1].abs()
l1_reg = self.alpha * l1_reg
penalties = self.root.get_penalty()
C = 0.
for (penalty, lmbda) in penalties:
C -= lmbda * 0.5 * (torch.log(penalty) + torch.log(1 - penalty))
output = torch.stack(max_Q)
self.root.reset() # reset all stacked calculation
return loss + C + l1_reg, output
# -log(loss) will always output non, because loss is always below zero.
# I suspect this is the mistake of the paper?
def collect_parameters(self):
nodes = [self.root]
self.module_list = nn.ModuleList()
self.param_list = nn.ParameterList()
while nodes:
node = nodes.pop(0)
if node.leaf:
param = node.param
self.param_list.append(param)
else:
fc = node.fc
# beta = node.beta
nodes.append(node.right)
nodes.append(node.left)
# self.param_list.append(beta)
self.module_list.append(fc)
def train_(self, train_loader, epoch):
self.train()
self.define_extras(self.args.batch_size)
for batch_idx, (data, target) in enumerate(train_loader):
correct = 0
if self.args.cuda:
data, target = data.cuda(), target.cuda()
target = Variable(target)
target_ = target.view(-1, 1)
batch_size = target_.size()[0]
data = data.view(batch_size, -1)
# convert int target to one-hot vector
data = Variable(data)
data = self.bn(data)
if not batch_size == self.args.batch_size: # because we have to initialize parameters for batch_size,
# tensor not matches with batch size cannot be trained
self.define_extras(batch_size)
self.target_onehot.data.zero_()
self.target_onehot.scatter_(1, target_, 1.)
self.optimizer.zero_grad()
loss, output = self.cal_loss(data, self.target_onehot)
# loss, output = self.cal_loss(data, target)
loss.backward(retain_variables=True)
# self.scheduler.step(loss.data[0])
self.optimizer.step()
pred = output.data.max(1)[1] # get the index of the max log-probability
correct += pred.eq(target.data).cpu().sum()
accuracy = 100. * correct / len(data)
if batch_idx % self.args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}, Accuracy: {}/{} ({:.4f}%)'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.data[0],
correct, len(data),
accuracy))
def test_(self, test_loader, epoch):
self.eval()
self.define_extras(self.args.batch_size)
correct = 0
auc_roc_data = list(), list()
for data, target in test_loader:
if self.args.cuda:
data, target = data.cuda(), target.cuda()
target = Variable(target)
target_ = target.view(-1, 1)
batch_size = target_.size()[0]
data = data.view(batch_size, -1)
# convert int target to one-hot vector
data = Variable(data)
data = self.bn(data)
if not batch_size == self.args.batch_size: # because we have to initialize parameters for batch_size,
# tensor not matches with batch size cannot be trained
self.define_extras(batch_size)
self.target_onehot.data.zero_()
self.target_onehot.scatter_(1, target_, 1.)
_, output = self.cal_loss(data, self.target_onehot)
pred = output.data.max(1)[1] # get the index of the max log-probability
correct += pred.eq(target.data).cpu().sum()
auc_roc_data[0].extend(target.data.cpu().numpy().reshape(-1))
auc_roc_data[1].extend(output.data[:, 1].cpu().numpy().reshape(-1))
accuracy = 100. * correct / len(test_loader.dataset)
auc_roc = roc_auc_score(*auc_roc_data)
with open('test_results.txt', 'a') as o:
print('Test set, epoch {}: Accuracy: {}/{} ({:.4f}%),\tAUC ROC: {:.3f}'.format(
epoch,
correct, len(test_loader.dataset),
accuracy,
auc_roc),
file=o
)
self.test_acc.append(accuracy)
# not compatible with lrs.ReduceLROnPlateau
# if auc_roc > self.best_auc_roc:
# self.save_best('./result')
# self.best_auc_roc = auc_roc
def save_best(self, path):
try:
os.makedirs('./result')
except:
print('directory ./result already exists')
with open(os.path.join(path, 'best_model.pkl'), 'wb') as output_file:
pickle.dump(self, output_file)