知乎问题代码

# -*- coding: utf-8 -*-
"""
Created on Sat May 19 18:44:40 2018

@author: John Kwok
"""

# import
import numpy as np
import tensorflow as tf
import GetDataUtil
# 数据读取及预处理
'''
定义超参
'''
BATCH_SIZE = 128 # 批大小
EPOCH = 5 # 训练EPOCH次数
HIDDEN_UNIT = 512 
KERNEL_SIZE = 3 # 通用卷集核大小
COVN_1_CHANNELS = 128 # 第一层卷积层的输出chennel个数
COVN_2_CHANNELS = 64 # 第二层卷积层的输出chennel个数
INPUT_CHENNELS = 1 # 输入图片的chennels个数
SCNN_KERNEL_LENGTH = 9 # SCNN 卷集核的宽度
PAD = SCNN_KERNEL_LENGTH - 1 # 切片后需要pad的个数
# 切片需要后需要PAD
PADDING = [[0, 0],
           [0,0],
           [int(PAD / 2),PAD - int(PAD/2)],       
           [0,0]]

'''
数据读取
'''
#X_train_origin, X_test_origin, y_train, y_test = GetDataUtil.getTrainTestSet(dataPath = "../DataSet_NPSave/JustifiedData.npy",test_size = 0.1)

X_train_origin, X_test_origin, y_train, y_test = GetDataUtil.getTrainTestSet(dataPath = "../DataSet_NPSave/RandomCrop_NPAWF_Noise_orgin_ACC_005_10000.npy",test_size = 0.1)

'''2. Data preprocessing'''
def DataPreprocess(data):
    print("Data Preprocessing,Please wait...")
    data[:,:3,:] = (data[:,:3,:] - np.mean(data[:,:3,:]))/np.std(data[:,:3,:])
    data[:,3:6,:] = (data[:,3:6,:] - np.mean(data[:,3:6,:]))/np.std(data[:,3:6,:])
    
    # 特征构造
    sin = np.sin(data * np.pi / 2)
    cos = np.cos(data * np.pi / 2)
    X_2 = np.power(data,2)
    X_3 = np.power(data,3)   
    ACC_All = np.sqrt((np.power(data[:,0,:],2)+
                      np.power(data[:,1,:],2)+
                      np.power(data[:,2,:],2))/3)[:,np.newaxis,:]    
    Ay_Gz = (data[:,1,:] * data[:,5,:])[:,np.newaxis,:]
    Ay_2_Gz = (np.power(data[:,1,:],2) * data[:,5,:])[:,np.newaxis,:]
    Ay_Gz_2 = (np.power(data[:,5,:],2) * data[:,1,:])[:,np.newaxis,:]
    Ax_Gy = (data[:,0,:] * data[:,4,:])[:,np.newaxis,:]
    Ax_2_Gy = (np.power(data[:,0,:],2) * data[:,4,:])[:,np.newaxis,:]
    Ax_Gy_2 = (np.power(data[:,4,:],2) * data[:,0,:])[:,np.newaxis,:]
    
    Ax_Ay_Az = (data[:,0,:]*data[:,1,:]*data[:,2,:])[:,np.newaxis,:]
    
    newData = np.concatenate((data,sin,cos,X_3,X_2,ACC_All,Ay_Gz,Ay_2_Gz,Ay_Gz_2,Ax_Gy,
                           Ax_2_Gy,Ax_Gy_2,Ax_Ay_Az),axis = 1)
    
    # data *= 255
    print(np.min(data))
    print(np.max(data))
    
    print("Finished!")
    return newData
#
X_train = DataPreprocess(X_train_origin)
X_test = DataPreprocess(X_test_origin)
#
#
data = X_train[:,:,:,np.newaxis]
label = y_train
#data = np.random.randn(1000,38,300,1) # NHWC
#label = np.random.randint(5,size = 1000)
print(data.shape)
print(label.shape)

'''
声明待训练参数
'''
# regularizers 也可以尝试0.01
weights = {
    'w_conv1':tf.get_variable(name = 'w_conv1',
                         shape = [2,
                                  2,
                                  INPUT_CHENNELS,
                                  COVN_1_CHANNELS],
                         initializer = tf.truncated_normal_initializer(0.0001),
                         regularizer = tf.keras.regularizers.l2(l=0.1)),
    'w_conv2':tf.get_variable(name = 'w_conv2',
                         shape = [KERNEL_SIZE,
                                  KERNEL_SIZE,
                                  COVN_1_CHANNELS,
                                  COVN_2_CHANNELS],
                         initializer = tf.truncated_normal_initializer(0.001),
                         regularizer = tf.keras.regularizers.l2(l=0.1))
}

biases = {
    'b_conv1': tf.get_variable(name = 'b_conv1',
                              shape= [COVN_1_CHANNELS],
                              initializer = tf.zeros_initializer(),
                              regularizer = tf.keras.regularizers.l2(l=0.01)),
    'b_conv2': tf.Variable(tf.zeros(COVN_2_CHANNELS),
                               name = "b_conv2")                     
}

tf.summary.histogram('w_conv1',weights['w_conv1'])
tf.summary.histogram('w_conv2',weights['w_conv2'])

tf.summary.histogram('b_conv1',biases['b_conv1'])
tf.summary.histogram('b_conv2',biases['b_conv2'])

# 定义网络
x = tf.placeholder(tf.float32,shape = (None,data.shape[1],data.shape[2],data.shape[3]))
y = tf.placeholder(tf.int64,shape = (None,1))
onehot_labels = tf.reshape(tf.one_hot(y,depth = 5),shape=(-1,5))

# 第一个卷积层
with tf.name_scope('covn_1'):
     conv_out_1 = tf.nn.conv2d(input = x,
                            filter = weights['w_conv1'],
                            strides = [1, 1, 1, 1],
                            padding = "SAME",
                            use_cudnn_on_gpu = True,
                            data_format = 'NHWC',
                            dilations = [1, 1, 1, 1],
                            name = 'conv_out')
     relu_out_1 = tf.nn.relu(tf.nn.bias_add(conv_out_1,biases['b_conv1']),
                                            name = 'relu_out')
     pooling_out_1 = tf.nn.max_pool(relu_out_1,
                                  ksize = [1,3,3,1],
                                  strides = [1,1,1,1],
                                  padding = "SAME",
                                  data_format='NHWC',
                                  name = 'pooling_out')

# 第二个卷积层
with tf.name_scope('covn_2'):
     conv_out_2 = tf.nn.conv2d(input = pooling_out_1,
                            filter = weights['w_conv2'],
                            strides = [1, 1, 1, 1],
                            padding = "SAME",
                            use_cudnn_on_gpu = True,
                            data_format = 'NHWC',
                            dilations = [1, 1, 1, 1],
                            name = 'conv_out')
     relu_out_2 = tf.nn.relu(tf.nn.bias_add(conv_out_2,biases['b_conv2']),
                                            name = 'relu_out')     
     pooling_out_2 = tf.nn.max_pool(relu_out_2,
                                  ksize = [1,3,3,1],
                                  strides = [1,1,1,1],
                                  padding = "SAME",
                                  data_format='NHWC',
                                  name = 'pooling_out')

with tf.name_scope('output'):

     fc_input = tf.layers.flatten(pooling_out_2,name='flatten')
     fc1_out = tf.layers.dense(fc_input,
                              1024,
                              activation=None,
                              use_bias=True,
                              kernel_initializer=tf.truncated_normal_initializer(0.01),
                              bias_initializer=tf.zeros_initializer(),
                              kernel_regularizer=tf.keras.regularizers.l2(l=0.01),
                              bias_regularizer=tf.keras.regularizers.l2(l=0.01),
                              name='fc_1')
     fc2_out = tf.layers.dense(fc1_out,
                              512,
                              activation=None,
                              use_bias=True,
                              kernel_initializer=tf.truncated_normal_initializer(0.01),
                              bias_initializer=tf.zeros_initializer(),
                              kernel_regularizer=tf.keras.regularizers.l2(l=0.01),
                              bias_regularizer=tf.keras.regularizers.l2(l=0.01),
                              name='fc_2')
     fc3_out = tf.layers.dense(fc2_out,
                              256,
                              activation=None,
                              use_bias=True,
                              kernel_initializer=tf.truncated_normal_initializer(0.01),
                              bias_initializer=tf.zeros_initializer(),
                              kernel_regularizer=tf.keras.regularizers.l2(l=0.01),
                              bias_regularizer=tf.keras.regularizers.l2(l=0.01),
                              name='fc_3')
     logits = tf.layers.dense(fc3_out,
                              5,
                              activation=None,
                              use_bias=True,
                              kernel_initializer=tf.truncated_normal_initializer(0.01),
                              bias_initializer=tf.zeros_initializer(),
                              kernel_regularizer=tf.keras.regularizers.l2(l=0.01),
                              bias_regularizer=tf.keras.regularizers.l2(l=0.01),
                              name='fc_output')

loss = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits_v2(labels = onehot_labels,
                                                                 logits = logits),name = 'loss')
optimizer = tf.train.AdamOptimizer(learning_rate = 0.001,name = 'adam')
train_op = optimizer.minimize(loss,name = 'train_op')

correct_pred = tf.equal(tf.argmax(logits,1),tf.argmax(onehot_labels,1),name = 'correct_pred')
c = tf.cast(correct_pred,tf.float32)
accuracy = tf.reduce_mean(c,name = 'accuracy')

init_op = tf.global_variables_initializer()
     
tf.summary.scalar('loss',loss)
tf.summary.scalar('accuracy',accuracy)

from sklearn.utils import shuffle
with tf.Session() as sess:
     writer = tf.summary.FileWriter('./log/scnn/')
     writer.add_graph(sess.graph)
     merge_all = tf.summary.merge_all()
     sess.run(init_op)
     sess.graph.finalize()
     step = 0
     for epoch in range(EPOCH):
          X,Label = shuffle(data,label,random_state=None)
          idx = 0
          while idx < X.shape[0]:
               if(idx+BATCH_SIZE>X.shape[0]):
                    x_batch = X[idx:]
                    y_batch = Label[idx:].reshape(-1,1)
               else:
                    x_batch = X[idx:idx+BATCH_SIZE]
                    y_batch = Label[idx:idx+BATCH_SIZE].reshape(-1,1)
               print("Step:"+str(step))
               pre, _, log = sess.run([accuracy,train_op,merge_all],feed_dict = {x:x_batch,y:y_batch})
               print(pre)
               writer.add_summary(log,step)
               idx += BATCH_SIZE
               step += 1
          l,acc = sess.run([loss,accuracy],feed_dict = {x:x_batch,y:y_batch})

          print("Epoch " + str(epoch+1) + ", Minibatch Loss= " + 
                "{:.4f}".format(l) + ", Training Accuracy= " + 
                "{:.3f}".format(acc))
     print("Optimization Finished!")