-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathmodel.py
191 lines (159 loc) · 9.92 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
import tensorflow as tf
import numpy as np
import pandas as pd
import time
from LRP_linear_layer import *
class LSTM(object):
def __init__(self, time_step, hidden_size, num_layers, learning_rate, batch_size, input_features):
tf.reset_default_graph()
self.time_step = time_step
self.hidden_size = hidden_size
self.num_layers = num_layers
self.learning_rate = learning_rate
self.batch_size = batch_size
self.input_features = input_features
self.multi_lstm_cell = tf.nn.rnn_cell.MultiRNNCell([tf.nn.rnn_cell.BasicLSTMCell(num_units=n) for n in self.hidden_size])
self.inputs = tf.placeholder(dtype=tf.float32, shape=[None, self.time_step, self.input_features])
self.targets = tf.placeholder(dtype=tf.float32, shape=[None])
self.states, self.states_c_h = tf.nn.dynamic_rnn(cell=self.multi_lstm_cell, inputs=self.inputs, dtype=tf.float32)
self.states_last_h = self.states[:, -1, :]
self.y_pre = tf.squeeze(tf.layers.dense(inputs=self.states_last_h, units=1, activation=None, use_bias=False), 1)
self.loss = tf.reduce_mean(tf.square(self.targets - self.y_pre))
optimizer = tf.train.RMSPropOptimizer(learning_rate=self.learning_rate) # 优化器
self.train_op = optimizer.minimize(self.loss)
self.saver = tf.train.Saver() # Saver对象声明
def get_a_cell(self):
return tf.nn.rnn_cell.LSTMCell(name='basic_lstm_cell', num_units=self.hidden_size)
def train(self, x_train, y_train, n_epochs, save_path, log_every_n, early_stop):
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
best = [100, 0] # 记录loss和epoch
earlystop = early_stop # 早停止条件，若连续earlystop都没有出现更小的loss，则训练停止并保存最佳模型
train_length = x_train.shape[0]
split_index = round(train_length * 0.8)
x_validation = x_train[split_index:]
y_validation = y_train[split_index:]
x_train = x_train[:split_index]
y_train = y_train[:split_index]
for epoch in range(n_epochs):
index = np.arange(x_train.shape[0])
np.random.shuffle(index) # 训练集打乱数据顺序
x_train = x_train[index]
y_train = y_train[index]
start = time.time()
for iteration in range(x_train.shape[0] // self.batch_size + 1):
if iteration != x_train.shape[0] // self.batch_size:
x_batch = x_train[self.batch_size * iteration:self.batch_size * (iteration + 1)]
# print('x_batch', x_batch.shape)
label_batch = y_train[self.batch_size * iteration:self.batch_size * (iteration + 1)]
elif iteration == x_train.shape[0] // self.batch_size & x_train.shape[0] % self.batch_size != 0:
x_batch = x_train[self.batch_size * iteration:]
label_batch = y_train[self.batch_size * iteration:]
elif iteration == x_train.shape[0] // self.batch_size & x_train.shape[0] % self.batch_size == 0:
continue
feed = {self.inputs: x_batch, self.targets: label_batch}
loss, _ = sess.run([self.loss, self.train_op], feed_dict=feed)
# cumulative_loss.append(loss)
end = time.time()
# 计算验证集上loss
feed_validation = {self.inputs: x_validation, self.targets: y_validation}
loss_validation, _ = sess.run([self.loss, self.train_op], feed_dict=feed_validation)
if loss_validation < best[0]:
best = [loss_validation, epoch]
self.saver.save(sess, save_path + '/model', global_step=epoch)
if epoch - best[1] > earlystop: # 早停止条件
break
if epoch % log_every_n == 0:
print('epoch: {}/{}... '.format(epoch, n_epochs),
'loss: {:.6f}... '.format(loss_validation),
'{:.4f} sec/epoch'.format((end - start)))
def test(self, x_test, y_test, checkpoint_path):
with tf.Session() as sess:
self.saver.restore(sess, tf.train.latest_checkpoint(checkpoint_path + '/'))
# states, stacked_states, stacked_outputs = sess.run([self.states, self.stacked_states, self.stacked_outputs], feed_dict={self.inputs: x_test})
# print(x_test.shape)
y_hat_un = sess.run(self.y_pre, feed_dict={self.inputs: x_test})
y_hat = y_hat_un
y_labels = y_test
return y_hat, y_labels
def lrp(self, x_test, checkpoint_path, eps, bias_factor):
with tf.Session() as sess:
self.saver.restore(sess, tf.train.latest_checkpoint(checkpoint_path + '/'))
states, states_c_h = sess.run([self.states, self.states_c_h], feed_dict={self.inputs: x_test})
y_hat_un = sess.run(self.y_pre, feed_dict={self.inputs: x_test})
# print('x_test:', x_test.shape)
# print('states.shape', states.shape)
# print('states_c_h.shape', states_c_h[0][0].shape)
ckpt = tf.train.get_checkpoint_state(checkpoint_path + '/')
# print('ckpt:', ckpt.model_checkpoint_path)
# 读取最新权重数据
reader = tf.train.NewCheckpointReader(ckpt.model_checkpoint_path)
# all_variables = reader.get_variable_to_shape_map()
# print('all_variables:', all_variables)
# 由于dynamic_rnn没有返回各时期的长期状态，要根据权重前向计算
W_rnn = reader.get_tensor('rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel')
# print(W_rnn.shape)
b_rnn = reader.get_tensor('rnn/multi_rnn_cell/cell_0/basic_lstm_cell/bias')
# print(b_rnn.shape)
W_dense = reader.get_tensor('dense/kernel')
# print(W_dense.shape)
T = x_test.shape[1] # 时间步
d = int(W_rnn.shape[-1] / 4) # 神经网络隐藏层单元个数 要修改
e = x_test.shape[2] # 特征个数
C = 1
# gate indices (assuming the gate ordering in the LSTM weights is i,g,f,o):
idx = np.hstack((np.arange(0, d), np.arange(2 * d, 4 * d))).astype(int) # indices of gates i,f,o together
idx_i, idx_g, idx_f, idx_o = np.arange(0, d), np.arange(d, 2 * d), np.arange(2 * d, 3 * d), np.arange(3 * d, 4 * d)
h_states = np.zeros((T + 1, d)) # 各时间步隐含状态
c_states = np.zeros((T + 1, d)) # 各时间步长期状态
gates_x_h = np.zeros((T, 4 * d))
gates_pre = np.zeros((T, 4 * d))
gates_act = np.zeros((T, 4 * d))
# for i in range(x_test.shape[0]):
contribution = pd.DataFrame()
for i in x_test.shape[0]:
temp_x_test = x_test[i]
# print(temp_x_test[0].shape)
# print(h_states[0].shape)
# 沿时间步展开
for t in range(T):
# xt, ht-1拼接
x_h_concat = np.concatenate([temp_x_test[t], h_states[t-1]])
gates_x_h[t] = np.dot(W_rnn.T, x_h_concat) # 4d * (e+d)和(e+d) * 1,得到4d*1
gates_pre[t] = gates_x_h[t] + b_rnn
gates_act[t, idx] = 1.0 / (1.0 + np.exp(- gates_pre[t, idx]))
gates_act[t, idx_g] = np.tanh(gates_pre[t, idx_g])
c_states[t] = gates_act[t, idx_f] * c_states[t - 1] + gates_act[t, idx_i] * gates_act[t, idx_g]
# h_states[t] = gates_act[t, idx_o] * np.tanh(c_states[t])
h_states[t] = states[i][t]
y_dense = np.dot(W_dense.T, h_states[T - 1])[0]
# print('隐藏状态：')
# print('原始：', states[0])
# print('前向计算：', h_states)
# print('长期状态：')
# print('原始：', states_c_h[0][0][0])
# print('前向计算：', c_states)
# print('预测值：')
# print('原始：', y_hat_un[0])
# print('前向计算：', y_dense)
# LRP
Rx = np.zeros(temp_x_test.shape)
Rh_Left = np.zeros((T + 1, d))
Rc_Left = np.zeros((T + 1, d))
Rg_Left = np.zeros((T, d)) # gate g only
Rout_mask = np.zeros((C))
Rout_mask[0] = 1.0
Rh_Left[T - 1] = lrp_linear(h_states[T - 1], W_dense, np.zeros((C)), np.array([y_hat_un[i]]), np.array([y_hat_un[i]]) * Rout_mask, d, eps, bias_factor, debug=False)
for t in reversed(range(T)):
Rc_Left[t] += Rh_Left[t]
Rc_Left[t - 1] = lrp_linear(gates_act[t, idx_f] * c_states[t - 1], np.identity(d), np.zeros((d)), c_states[t], Rc_Left[t], d, eps, bias_factor, debug=False)
Rg_Left[t] = lrp_linear(gates_act[t, idx_i] * gates_act[t, idx_g], np.identity(d), np.zeros((d)), c_states[t], Rc_Left[t], d, eps, bias_factor, debug=False)
# self.Wxh_Left[idx_g].T
Rx[t] = lrp_linear(temp_x_test[t], W_rnn[np.arange(0, e)].T[idx_g].T, b_rnn[idx_g], gates_pre[t, idx_g], Rg_Left[t], d + e, eps, bias_factor, debug=False)
Rh_Left[t - 1] = lrp_linear(h_states[t - 1], W_rnn[np.arange(e, e + d)].T[idx_g].T, b_rnn[idx_g], gates_pre[t, idx_g], Rg_Left[t], d + e, eps, bias_factor, debug=False)
# print('Rx_i.shape:', Rx_i.shape)
contribution_i = np.sum(Rx, axis=0)
contribution_i_df = pd.DataFrame(contribution_i)
# print(contribution_i_df)
contribution = pd.concat([contribution, contribution_i_df], axis=1, sort=True)
return contribution