DNN.py_Pythondnn_机器学习_DNN神经网络_DNN_

from __future__ import print_function import tensorflow as tf import numpy as np import scipy.io as sio import time import random import h5py # Functions for deep neural network weights initialization def ini_weights(n_input, n_hidden_1, n_hidden_2, n_hidden_3, n_output): weights = { 'h1': tf.Variable(tf.truncated_normal([n_input, n_hidden_1]) / np.sqrt(n_input)), 'h2': tf.Variable(tf.truncated_normal([n_hidden_1, n_hidden_2]) / np.sqrt(n_hidden_1)), 'h3': tf.Variable(tf.truncated_normal([n_hidden_2, n_hidden_3]) / np.sqrt(n_hidden_2)), 'out': tf.Variable(tf.truncated_normal([n_hidden_3, n_output])) / n_hidden_3, } biases = { 'b1': tf.Variable(tf.ones([n_hidden_1]) * 0.1), 'b2': tf.Variable(tf.ones([n_hidden_2]) * 0.1), 'b3': tf.Variable(tf.ones([n_hidden_3]) * 0.1), 'out': tf.Variable(tf.ones([n_output]) * 0.1), } return weights, biases # Functions for deep neural network structure construction def multilayer_perceptron(x, weights, biases, input_keep_prob, hidden_keep_prob): x = tf.nn.dropout(x, input_keep_prob) # dropout layer layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1']) # x = wx+b layer_1 = tf.nn.relu(layer_1) # x = max(0, x) layer_1 = tf.nn.dropout(layer_1, hidden_keep_prob) # dropout layer layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2']) layer_2 = tf.nn.relu(layer_2) layer_2 = tf.nn.dropout(layer_2, hidden_keep_prob) layer_3 = tf.add(tf.matmul(layer_2, weights['h3']), biases['b3']) layer_3 = tf.nn.relu(layer_3) layer_3 = tf.nn.dropout(layer_3, hidden_keep_prob) out_layer = tf.matmul(layer_3, weights['out']) + biases['out'] out_layer = tf.nn.relu(out_layer) return out_layer # Functions for deep neural network training def train(X, Y, location, training_epochs=10, batch_size=100, LR=0.0001, n_hidden_1=1000, n_hidden_2=800, n_hidden_3=1500, traintestsplit=0.2, LRdecay=1): num_total = X.shape[0] # number of total samples num_val = int(num_total * traintestsplit) # number of validation samples num_train = num_total - num_val # number of training samples n_input = X.shape[1] # input size n_output = Y.shape[1] # output size X_train = X[0:num_train, :] # training data Y_train = Y[0:num_train, :] # training label X_val = X[num_train:num_total, :] # validation data Y_val = Y[num_train:num_total, :] # validation label # vData = h5py.File('/home/b253/Desktop/fsy2/31second.mat') # vData_t = np.transpose(vData['f2']) # np.save('31second.npy', vData_t) # vaData = np.load('31second.npy') # vnum_all = len(vaData) # vindex = [i for i in range(vnum_all)] # i=0~10679 # random.shuffle(vindex) # 实现编号的乱序 # vaData = vaData[vindex] # X_val =vaData[0:vnum_all, 5:1005] # validation data # Y_val =vaData[0:vnum_all, 0:5] # validation label x = tf.placeholder("float", [None, n_input]) y = tf.placeholder("float", [None, n_output]) is_train = tf.placeholder("bool") learning_rate = tf.placeholder(tf.float32, shape=[]) total_batch = int(num_total / batch_size) print('train: %d ' % num_train,'test_accuracy: %d ' % num_val) input_keep_prob = tf.placeholder(tf.float32) hidden_keep_prob = tf.placeholder(tf.float32) weights, biases = ini_weights(n_input, n_hidden_1, n_hidden_2, n_hidden_3, n_output) pred = multilayer_perceptron(x, weights, biases, input_keep_prob, hidden_keep_prob) cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=pred, labels=tf.argmax(y, 1))) optimizer = tf.train.RMSPropOptimizer(learning_rate, 0.9).minimize(cost) # training algorithms: RMSprop correct_predection = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1)) accuracy = tf.reduce_mean(tf.cast(correct_predection, tf.float32)) init = tf.global_variables_initializer() saver = tf.train.Saver() #声明tf.train.Saver类用于保存模型 MSETime = np.zeros((training_epochs, 3)) with tf.Session() as sess: sess.run(init) start_time = time.time() for epoch in range(training_epochs):#训练迭代10轮 for i in range(total_batch): idx = np.random.randint(num_train, size=batch_size)#每一个批次训练就从训练数目中随机生成200个索引编号 if LRdecay == 1: _, c = sess.run([optimizer, accuracy], feed_dict={x: X_train[idx, :], y: Y_train[idx, :], input_keep_prob: 1, hidden_keep_prob: 1, learning_rate: LR / (epoch + 1), is_train: True}) elif LRdecay == 0: _, c = sess.run([optimizer, accuracy], feed_dict={x: X_train[idx, :], y: Y_train[idx, :], input_keep_prob: 1, hidden_keep_prob: 1, learning_rate: LR, is_train: True}) MSETime[epoch, 0] = c MSETime[epoch, 1] = sess.run(accuracy, feed_dict={x: X_val, y: Y_val, input_keep_prob: 1, hidden_keep_prob: 1, is_train: False}) MSETime[epoch, 2] = time.time() - start_time if epoch % (int(training_epochs / 10)) == 0: #改 print('epoch:%d, ' % epoch, 'test_accuracy:%0.2f%%.' % (c * 100)) print("training time: %0.2f s" % (time.time() - start_time)) sio.savemat('MSETime_%d_%d_%d' % (n_output, batch_size, LR * 10000), {'train': MSETime[:, 0], 'validation': MSETime[:, 1], 'time': MSETime[:, 2]}) saver.save(sess, location) #将模型保存到location中的model_demo.ckpt文件中 # sio.savemat(save_name, {'pred': pred, 'Y': Y_train[idx, :]}) # 预测值来自训练集和验证集，因此前面的返回值一定要指定 return 0 # Functions for deep neural network testing def test(X, Y, model_location, save_name, n_input, n_output, n_hidden_1=1000, n_hidden_2=800, n_hidden_3=1500): tf.reset_default_graph() x = tf.placeholder("float", [None, n_input]) y = tf.placeholder("float", [None, n_output]) is_train = tf.placeholder("bool") input_keep_prob = tf.placeholder(tf.float32) hidden_keep_prob = tf.placeholder(tf.float32) weights, biases = ini_weights(n_input, n_hidden_1, n_hidden_2, n_hidden_3, n_output) pred = multilayer_perceptron(x, weights, biases, input_keep_prob, hidden_keep_prob) correct_predection = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1)) accuracy = tf.reduce_mean(tf.cast(correct_predection, tf.float32)) saver = tf.train.Saver() #加载模型之前，同样要声明模型保存的类 with tf.Session() as sess: #不再运行变量的初始化过程 saver.restore(sess, model_location) #加载已经保存的模型,并通过模型中保存的变量值来计算预测值 start_time = time.time() y_pred, test_accuracy = sess.run([pred, accuracy], feed_dict={x: X, y: Y, input_keep_prob: 1, hidden_keep_prob: 1, is_train: False}) testtime = time.time() - start_time # print("testing time: %0.2f s" % testtime) print("test_accuracy is %g" % test_accuracy) sio.savemat(save_name, {'pred': y_pred, 'Y': Y}) return dnntime