def making2(p, argchoice_r):
'''
选择特殊半径组成数据集制作数据
:param p: 数据集路径, str
:param argchoice_r: 半径列表, list
:return: 子数据集
'''
dataset = LoadFile(p)
dataset_pd = pd.DataFrame(
data=dataset,
columns=['f' + '%s' % i
for i in range(dataset.shape[-1] - 1)] + ['label'])
rs = dataset[:, -1]
rs_dict = Counter(rs)
rs_list = list(rs_dict.keys())
rs_list = sorted(rs_list)
choice_r = list(filter(lambda x: rs_list.index(x) in argchoice_r, rs_list))
#统计所有半径值数据中数据量最大的数据量值
max = 0
for i in choice_r:
data_subpd = dataset_pd.loc[dataset_pd['label'] == i]
if data_subpd.values.shape[0] > max:
max = data_subpd.values.shape[0]
data_all = np.array([0])
data_array_con = np.zeros(dtype=np.float32, shape=[1])
for i in choice_r:
data_subpd = dataset_pd.loc[dataset_pd['label'] == i]
# print(data_subpd.values.shape)
num = max // data_subpd.values.shape[0]
num = num if max % data_subpd.values.shape[0] == 0 else num + 1
data_array = data_subpd.values
np.random.shuffle(data_array)
for i in range(num):
if data_array_con.any() == 0:
data_array_con = data_array
else:
data_array_con = np.vstack((data_array_con, data_array))
print(data_array_con.shape, max)
data_all = data_array_con[:max, :] if data_all.any(
) == 0 else np.vstack((data_all, data_array_con[:max, :]))
data_array_con = np.array([0])
return data_all
def session(dataset_path, train_path='', test_path=''):
'''
节点连接
:param dataset_path: 数据集路径
:param train_path: 训练集数据路径,默认为空
:param test_path: 测试集数据路径,默认为空
:return: None
'''
#导入数据集
dataset = LoadFile(p=dataset_path)
# dataset = guiyi(dataset)
dataset = onehot(dataset)
g = tf.Graph()
with g.as_default():
with tf.name_scope('placeholder'):
x_f = tf.placeholder(dtype=tf.float32, shape=[None, 4], name='x_f')
x_l = tf.placeholder(dtype=tf.float32,
shape=[None, 100],
name='x_l')
y = tf.placeholder(dtype=tf.float32, shape=[None, 3], name='y')
learning_rate = tf.placeholder(dtype=tf.float32, name='lr')
is_training = tf.placeholder(dtype=tf.bool, name='is_training')
output = layers(x_f=x_f, x_l=x_l, is_training=is_training)
with tf.name_scope('prediction'):
# loss = -tf.reduce_mean(y * tf.log(output), name='loss')
loss = tf.reduce_mean(
tf.keras.losses.categorical_crossentropy(y_true=y,
y_pred=output))
opt = tf.train.GradientDescentOptimizer(
learning_rate=learning_rate).minimize(loss)
acc = tf.reduce_mean(tf.cast(
tf.equal(tf.keras.backend.argmax(output, axis=1),
tf.keras.backend.argmax(y, axis=1)), tf.float32),
name='pred')
with tf.name_scope('etc'):
init = tf.global_variables_initializer()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options),
graph=g) as sess:
sess.run(init)
#划分训练集和测试集
train_data, test_data = spliting(dataset, 3000)
for i in range(6000): #6000
for data in input(dataset=train_data, batch_size=500):
_ = sess.run(opt,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-3],
y: data[:, -3:],
learning_rate: 1e-2,
is_training: False
})
if i % 100 == 0:
loss_ = sess.run(loss,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-3],
y: data[:, -3:],
is_training: False
})
acc_1 = sess.run(acc,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-3],
y: data[:, -3:],
is_training: False
})
if i % 100 == 0:
acc_2 = sess.run(acc,
feed_dict={
x_f: test_data[:, :4],
x_l: test_data[:, 4:-3],
y: test_data[:, -3:],
is_training: False
})
print('第%s轮训练集损失函数值为: %s 训练集准确率为: %s 测试集准确率为: %s' %
(i, loss_, acc_1, acc_2))
tf.summary.FileWriter('log/first_graph', sess.graph)
input_secondc2 = input_1.loc[input_1['%s' % input.shape[-1]] == 2]
input_secondc2 = input_secondc2.drop(columns=['%s' % input.shape[-1]])
input_secondc2 = input_secondc2.values
#放入11分类器中进一步分类
r_finally3 = np.array(second_check2(input_secondc2))[:, np.newaxis]
r_finally = np.vstack((r_finally1, r_finally2, r_finally3))
r_real = np.vstack(
(input_firstc[:, -1][:, np.newaxis], input_secondc1[:, -1][:,
np.newaxis],
input_secondc2[:, -1][:, np.newaxis]))
return r_finally, r_real
if __name__ == '__main__':
p = r'/home/xiaosong/桌面/pny相关数据/data_pny/PNY_all.pickle'
input = LoadFile(p=p)
np.random.shuffle(input)
dataset_4feature, dataset_dense, label = input[:, :
4], input[:, 4:
-1], input[:,
-1][:,
np.
newaxis]
dataset_fft = fft_transformer(dataset_dense, 100)
dataset = np.hstack((dataset_4feature, dataset_fft, label))
dataset_guiyi_2 = guiyi(dataset)
print(dataset_guiyi_2.shape)
r_finally, r_real = check(input=dataset_guiyi_2[:5100, :])
r_1 = np.where(np.abs(r_finally - r_real) < 1e-2, 1, 0)
r_sum = np.sum(r_1)
acc = r_sum / r_1.shape[0]
def session(dataset_path, train_path='', test_path=''):
'''
节点连接
:param dataset_path: 数据集路径
:param train_path: 训练集数据路径,默认为空
:param test_path: 测试集数据路径,默认为空
:return: None
'''
#导入数据集
dataset = LoadFile(p=dataset_path)
g1 = tf.Graph()
with g1.as_default():
with tf.name_scope('placeholder'):
x_f = tf.placeholder(dtype=tf.float32, shape=[None, 4], name='x_f')
x_l = tf.placeholder(dtype=tf.float32,
shape=[None, 100],
name='x_l')
y = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='y')
learning_rate = tf.placeholder(dtype=tf.float32, name='lr')
is_training = tf.placeholder(dtype=tf.bool, name='is_training')
# output = res_regression(x_f=x_f, x_l=x_l, is_training=is_training)
output = cnnlstm_regression(x_f=x_f, x_l=x_l, is_training=is_training)
with tf.name_scope('prediction'):
loss = tf.reduce_mean(tf.square(output - y))
opt = tf.train.GradientDescentOptimizer(
learning_rate=learning_rate).minimize(loss)
# opt = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss)
acc = acc_regression(Threshold=0.1, y_true=y, y_pred=output)
with tf.name_scope('etc'):
init = tf.global_variables_initializer()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options),
graph=g1) as sess:
sess.run(init)
# 划分训练集和测试集
train_data, test_data = spliting(dataset, 2269)
for i in range(2000): #20000
for data in input(dataset=train_data, batch_size=1000):
_ = sess.run(opt,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
learning_rate: 1e-2,
is_training: False
})
if i % 100 == 0:
loss_ = sess.run(loss,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
acc_1 = sess.run(acc,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
if i % 100 == 0:
acc_2 = sess.run(acc,
feed_dict={
x_f: test_data[:, :4],
x_l: test_data[:, 4:-1],
y: test_data[:, -1][:, np.newaxis],
is_training: False
})
print('第%s轮训练集损失函数值为: %s 训练集准确率为: %s 测试集准确率为: %s' %
(i, loss_, acc_1, acc_2))
def session(dataset_path, train_path='', test_path=''):
'''
节点连接
:param dataset_path: 数据集路径
:param train_path: 训练集数据路径,默认为空
:param test_path: 测试集数据路径,默认为空
:return: None
'''
#导入数据集
dataset = LoadFile(p=dataset_path)
g1 = tf.Graph()
with g1.as_default():
with tf.name_scope('placeholder'):
x_f = tf.placeholder(dtype=tf.float32, shape=[None, 4], name='x_f')
x_l = tf.placeholder(dtype=tf.float32,
shape=[None, 100],
name='x_l')
y = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='y')
learning_rate = tf.placeholder(dtype=tf.float32, name='lr')
is_training = tf.placeholder(dtype=tf.bool, name='is_training')
# output = res_regression(x_f=x_f, x_l=x_l, is_training=is_training)
output = cnnlstm_regression(x_f=x_f, x_l=x_l, is_training=is_training)
with tf.name_scope('prediction'):
loss = tf.reduce_mean(tf.square(output - y))
opt = tf.train.GradientDescentOptimizer(
learning_rate=learning_rate).minimize(loss)
# opt = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss)
acc1 = acc_regression(Threshold=0.1, y_true=y, y_pred=output)
acc2 = acc_regression(Threshold=0.16, y_true=y, y_pred=output)
acc3 = acc_regression(Threshold=0.2, y_true=y, y_pred=output)
acc4 = acc_regression(Threshold=0.3, y_true=y, y_pred=output)
with tf.name_scope('etc'):
init = tf.global_variables_initializer()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options),
graph=g1) as sess:
sess.run(init)
# 划分训练集和测试集
train_data, test_data = spliting(dataset, 3369)
for i in range(60000): #20000
for data in input(dataset=train_data, batch_size=1000):
_ = sess.run(opt,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
learning_rate: 1e-2,
is_training: False
})
if i % 100 == 0:
loss_ = sess.run(loss,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
acc_1 = sess.run(acc1,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
acc_2 = sess.run(acc2,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
acc_3 = sess.run(acc3,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
acc_4 = sess.run(acc4,
feed_dict={
x_f: data[:, :4],
x_l: data[:, 4:-1],
y: data[:, -1][:, np.newaxis],
is_training: False
})
if i % 100 == 0:
acc_5 = sess.run(acc1,
feed_dict={
x_f: test_data[:, :4],
x_l: test_data[:, 4:-1],
y: test_data[:, -1][:, np.newaxis],
is_training: False
})
acc_6 = sess.run(acc2,
feed_dict={
x_f: test_data[:, :4],
x_l: test_data[:, 4:-1],
y: test_data[:, -1][:, np.newaxis],
is_training: False
})
acc_7 = sess.run(acc3,
feed_dict={
x_f: test_data[:, :4],
x_l: test_data[:, 4:-1],
y: test_data[:, -1][:, np.newaxis],
is_training: False
})
acc_8 = sess.run(acc4,
feed_dict={
x_f: test_data[:, :4],
x_l: test_data[:, 4:-1],
y: test_data[:, -1][:, np.newaxis],
is_training: False
})
print(
'第%s轮训练集损失函数值为: %s 训练集准确率为: %s:%s %s:%s %s:%s %s:%s 测试集准确率为: %s:%s %s:%s %s:%s %s:%s'
% (i, loss_, 0.1, acc_1, 0.16, acc_2, 0.2, acc_3, 0.3,
acc_4, 0.1, acc_5, 0.16, acc_6, 0.2, acc_7, 0.3, acc_8))
tf.summary.FileWriter('log/regression_cnnlstm', sess.graph)
# 保存模型到文件当前脚本文件路径下的pb格式
saving_model = SaveImport_model(
sess_ori=sess,
file_suffix=r'/cnnlstm',
ops=(output, x_f, x_l, is_training),
usefulplaceholder_count=4,
pb_file_path=r'/home/xiaosong/桌面/regression_cnnlstm')
saving_model.save_pb()
def fft_transformer(dataset, N):
'''
对矩阵中各行按照指定点数做FFT变换
:param dataset: 待处理矩阵
:param N: 变换后点数
:return: 处理后矩阵
'''
fft_abs = np.abs(np.fft.fft(a=dataset, n=N, axis=1))
return fft_abs
if __name__ == '__main__':
p = r'/home/xiaosong/桌面/OLDENBURG_all.pickle'
dataset = LoadFile(p)
nums_cl = [[6557, 0], [611, 2], [101, 2], [13, 2], [554, 2], [155, 2],
[100, 2], [1165, 1], [1993, 1], [947, 2], [1133, 2], [1152, 1],
[542, 2], [754, 2], [2163, 1]]
dataset_output = making(nums_cl=nums_cl, dataset=dataset)
print(dataset_output.shape)
checkclassifier(dataset_output[:, -1])
# SaveFile(dataset_output, savepickle_p=r'/home/xiaosong/桌面/OLDENBURG_3cl.pickle')
dataset_4feature, dataset_dense, label = dataset_output[:, :
4], dataset_output[:,
4:
-1], dataset_output[:,
-1][:,
np
.
newaxis]