LSTM（长短期记忆神经网络）的小项目

""" LSTM-长短期记忆网络 Datetime：2020-9-8 purpose:预测20世纪70年代中期波士顿郊区房屋价格的中位数。使用boston_housing数据集 Data:404个训练样本和102个测试样本，包含地区的犯罪率、当地房产税率等13个特征量 Author：kris.wang """ from keras.models import Sequential from keras.layers import Dense,Dropout,Activation from keras.layers import Embedding from keras.layers import SimpleRNN,LSTM,GRU from keras.optimizers import SGD,Nadam,Adam,RMSprop from keras.callbacks import TensorBoard from keras.utils import np_utils import scipy.io as si import os import numpy as np import matplotlib.pyplot as plt #1,导入数据 # path=os.getcwd().replace('smallprojectsummary\LSTMproject','') data=si.loadmat('sp1s_aa_1000Hz.mat') # y_test=np.loadtxt(path+r'BCI_II_data\') #预处理数据 #316个样本，28个电极通道，500毫秒数据。 data_train=data['x_train'].reshape((316,500,28)) data_train/=200 data_train=data_train.astype('float32') data_test=data['x_test'].reshape((100,500,28)) data_test/=200 data_test=data_test.astype('float32') data_label=data['y_train'].reshape((316,1)) tmp_train = [] for i in data_label: if i == 1: tmp_train.append(1) elif i == 0: tmp_train.append(-1) data_label = np.array(tmp_train) data_label = np_utils.to_categorical(data_label, 2) data_label = data_label.astype('float32') # print(data_label.shape) #训练模型 def build_model(): model=Sequential() model.add(LSTM(10,return_sequences=True,input_shape=(500,28))) model.add(LSTM(10,return_sequences=True)) model.add(LSTM(5)) model.add(Dense(2,activation='softmax')) model.summary() model.compile( optimizer=Nadam(lr=0.001), loss='categorical_crossentropy', metrics=['acc'] ) # model.summary()#显示模型结构。 return model # score,acc=model.evaluate() # k折交叉验证 k=4 num_samples=len(data_train)//k all_acc_histories=[] for i in range(k): print('processing fold #',i) #准备验证数据 val_data=data_train[i*num_samples:(i+1)*num_samples] val_targets=data_label[i*num_samples:(i+1)*num_samples] #准备训练数据 partial_train_data=np.concatenate( [data_train[:i*num_samples], data_train[(i+1)*num_samples:]], axis=0 ) partial_train_targets=np.concatenate( [data_label[:i*num_samples], data_label[(i+1)*num_samples:]], axis=0 ) #构建模型 model=build_model() history=model.fit(partial_train_data,partial_train_targets,validation_data=(val_data,val_targets),epochs=80,batch_size=20)#verbose静默模式 val_los,val_acc=model.evaluate(val_data,val_targets,verbose=0)#给出损失值和目标值 #记录每个轮次中所有折mae acc_history=history.history['val_acc'] all_acc_histories.append(acc_history)#验证分数，即mae #计算所有轮次中的K折验证分数平均值 average_acc_history=[np.mean([x[i] for x in all_acc_histories]) for i in range(80)] #绘制验证分数 def smooth_curve(points,factor=0.9): smoothed_points=[] for point in points: if smoothed_points: previous=smoothed_points[-1] smoothed_points.append(previous*factor+point*(1-factor)) else: smoothed_points.append(point) return smoothed_points if __name__=="__main__": smooth_mae_history = smooth_curve(average_acc_history[10:]) # 因为前10个数据点的取值范围和其它点不同，所以删除它们。 plt.plot(range(1, len(smooth_mae_history) + 1), smooth_mae_history) plt.xlabel('Epochs') plt.ylabel('Validation MAE') plt.show() # -从图中能明显发现模型在80轮次后出现了过拟合。