结构 5 层卷积 - 3 层全连接 使用 SVM 取代 softmax 进行预测; 计算量有点大,大家看看即可。 卷积网络结构可以参考 AlexNet
%%time
import numpy as np
import matplotlib.pylab as plt
%matplotlib inline
import tensorflow as tf
from sklearn.cross_validation import train_test_split
fac = np.load('F:/Quotes/fac16.npy').astype(np.float32)
ret = np.load('F:/Quotes/ret16.npy').astype(np.float32)
train_X, test_X, train_Y, test_Y = train_test_split(fac, ret, test_size= 0.4)
print ('训练集 /总数据集 %.3f'%(len(train_X)/len(fac)))
# Parameters
learning_rate = 0.001 # 学习速率,
training_iters = 20 # 训练次数
batch_size = 1024 # 每次计算数量 批次大小
display_step = 10 # 显示步长
# Network Parameters
n_input = 40*17 # 40 天×17 多因子
n_classes = 7 # 根据涨跌幅度分成 7 类别
# 这里注意要使用 one-hot 格式,也就是如果分类如 3 类 -1,0,1 则需要 3 列来表达这个分类结果, 3 类是-1 0 1 然后是哪类,哪类那一行为 1 否则为 0
dropout = 0.5# Dropout, probability to keep units
# tensorflow 图 Graph 输入 input ,这里的占位符均为输入
x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])
keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)
# 2 层 CNN 提取特征向量
def CNN_Net_two(x,weights,biases,dropout=0.8,m=1):
# layer hidden 1
x = tf.reshape(x, shape=[-1,40,17,1])
x = tf.nn.conv2d(x, weights['wc1'], strides=[1,m,m,1],padding='SAME')
x = tf.nn.bias_add(x,biases['bc1'])
x = tf.nn.relu(x)
x = tf.nn.local_response_normalization(x, depth_radius=5, bias=1.0, alpha=0.001/9.0)
x = tf.nn.dropout(x,0.3)
# layer hidden 2
x = tf.nn.conv2d(x, weights['wc2'], strides=[1,m,m,1],padding='SAME')
x = tf.nn.bias_add(x,biases['bc2'])
x = tf.nn.relu(x)
x = tf.nn.local_response_normalization(x, depth_radius=5, bias=1.0, alpha=0.001/9.0)
x = tf.nn.dropout(x,0.3)
# layer hidden 3
x = tf.nn.conv2d(x, weights['wc3'], strides=[1,m,m,1],padding='SAME')
x = tf.nn.bias_add(x,biases['bc3'])
x = tf.nn.relu(x)
x = tf.nn.local_response_normalization(x, depth_radius=5, bias=1.0, alpha=0.001/9.0)
x = tf.nn.dropout(x,0.3)
# layer hidden 4
x = tf.nn.conv2d(x, weights['wc4'], strides=[1,m,m,1],padding='SAME')
x = tf.nn.bias_add(x,biases['bc4'])
x = tf.nn.relu(x)
x = tf.nn.local_response_normalization(x, depth_radius=5, bias=1.0, alpha=0.001/9.0)
x = tf.nn.dropout(x,0.3)
# layer hidden 5
x = tf.nn.conv2d(x, weights['wc5'], strides=[1,m,m,1],padding='SAME')
x = tf.nn.bias_add(x,biases['bc5'])
x = tf.nn.relu(x)
x = tf.nn.local_response_normalization(x, depth_radius=5, bias=1.0, alpha=0.001/9.0)
x = tf.nn.dropout(x,0.3)
#print (x.get_shape().as_list())
# 全连接层 1
x = tf.reshape(x,[-1,weights['wd1'].get_shape().as_list()[0]])
x = tf.add(tf.matmul(x,weights['wd1']),biases['bd1'])
x = tf.nn.relu(x)
x = tf.nn.dropout(x,dropout)
#print (x.get_shape().as_list())
# 全连接层 2
x = tf.reshape(x,[-1,weights['wd2'].get_shape().as_list()[0]])
x = tf.add(tf.matmul(x,weights['wd2']),biases['bd2'])
x = tf.nn.relu(x)
x = tf.nn.dropout(x,dropout)
#print (x.get_shape().as_list())
# 全连接层 3
x = tf.reshape(x,[-1,weights['wd3'].get_shape().as_list()[0]])
x = tf.add(tf.matmul(x,weights['wd3']),biases['bd3'])
x = tf.nn.relu(x)
x = tf.nn.dropout(x,dropout)
#print (x.get_shape().as_list())
t = tf.add(tf.matmul(x,weights['out']),biases['out'])
#print (t.get_shape().as_list())
# 返回两个数值, t 用于 softmax 分类, x 用于提取 CNN 处理的数据,也就是经过卷积处理的特征向量。
return t,x
# Store layers weight & bias
weights = {
'wc1': tf.Variable(tf.random_normal([10, 5, 1, 64])),
'wc2': tf.Variable(tf.random_normal([10, 5, 64, 128])),
'wc3': tf.Variable(tf.random_normal([10, 5, 128, 256])),
'wc4': tf.Variable(tf.random_normal([10, 5, 256, 512])),
'wc5': tf.Variable(tf.random_normal([10, 5, 512, 1024])),
'wd1': tf.Variable(tf.random_normal([40*17*1024, 1024])),
'wd2': tf.Variable(tf.random_normal([1024, 256])),
'wd3': tf.Variable(tf.random_normal([256, 32])),
'out': tf.Variable(tf.random_normal([32, n_classes]))
}
biases = {
'bc1': tf.Variable(tf.random_normal([64])),
'bc2': tf.Variable(tf.random_normal([128])),
'bc3': tf.Variable(tf.random_normal([256])),
'bc4': tf.Variable(tf.random_normal([512])),
'bc5': tf.Variable(tf.random_normal([1024])),
'bd1': tf.Variable(tf.random_normal([1024])),
'bd2': tf.Variable(tf.random_normal([256])),
'bd3': tf.Variable(tf.random_normal([32])),
'out': tf.Variable(tf.random_normal([n_classes]))
}
# 模型优化
pred,tmp = CNN_Net_two(x,weights,biases,dropout=keep_prob)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred,y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
correct_pred = tf.equal(tf.argmax(pred,1),tf.arg_max(y,1))
# tf.argmax(input,axis=None) 由于标签的数据格式是 -1 0 1 3 列,该语句是表示返回值最大也就是 1 的索引,两个索引相同则是预测正确。
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
# 更改数据格式,降低均值
init = tf.global_variables_initializer()
计算保存模型
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
# for step in range(300):
for step in range(1):
trl=int(len(train_X)/batch_size)
for i in range(trl):
print (i,'--',trl)
batch_x = train_X[i*batch_size:(i+1)*batch_size]
batch_y = train_Y[i*batch_size:(i+1)*batch_size]
sess.run(optimizer,feed_dict={x:batch_x,y:batch_y,keep_prob:0.5})
loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x,y: batch_y,keep_prob: 1.})
print("Iter " + str(step*batch_size) + ", Minibatch Loss= " + \
"{:.6f}".format(loss) + ", Training Accuracy= " + \
"{:.5f}".format(acc))
save_path = saver.save(sess,'F:/Quotes/test_var.ckpt')
print ('保持变量')
print("Optimization Finished!")
sess.close()
读取模型,进行预测
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
saver.restore(sess,'F:/Quotes/test_var.ckpt')
trainX_Convolution = sess.run(tmp, feed_dict={x:train_X, keep_prob:1.})
# 经过卷积处理的特征向量
nn_score = sess.run(accuracy,feed_dict={x:train_X, keep_prob:1.})
nn_score1 = sess.run(accuracy,feed_dict={x:test_X, keep_prob:1.})
print(nn_score,'---',nn_score1)
sess.close()
one-hot 向量转换为列向量
# train_Y
ol_train_Y = []
for i in range(len(train_Y)):
t = train_Y[i]
arg = np.argmax(t)
ol_train_Y.append(arg)
# softmax_pred
ol_softmax_pred = []
for i in range(len(softmax_pred)):
t = softmax_pred [i]
arg = np.argmax(t)
ol_softmax_pred.append(arg)
SVM 预测
from sklearn.svm import SVC
clf = SVC(C=0.9,gamma=1.0,decision_function_shape='ovo')
clf.fit(trainX_Convolution, ol_train_Y)
c = clf.predict(trainX_Convolution)
print ('CNN 预测',(np.corrcoef(a,c)[0][1]))
集成算法比较参见: https://uqer.io/community/share/58562a9f6a5e6d0052291ebe