利用keras搭建自编码器网络——脑机接口

因为我的本科毕设就是做自编码器的，所以，自编码器部分还是很好理解的，在这里不多赘述。

我现在主要是看一下mnist的自编码器部分，然后利用这个来搭建自己数据的自编器网络。

贴几个参考代码的链接

https://kiseliu.github.io/2016/08/16/building-autoencoders-in-keras/

http://ff120.github.io/2017/04/20/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0%E4%B8%93%E9%A2%98/%E4%BD%BF%E7%94%A8keras%E5%AE%9E%E7%8E%B0autoencoder/

https://keras-cn.readthedocs.io/en/latest/blog/autoencoder/

https://morvanzhou.github.io/tutorials/machine-learning/keras/3-1-save/

#!/usr/bin/python# -*- coding:utf8 -*-import kerasimport numpy as npfrom keras.datasets import mnistfrom keras.models import Modelfrom keras.layers import Dense, Inputfrom keras import regularizersimport matplotlib.pyplot as pltbatch_size = 256epochs = 100#数据预处理(x_train, _), (x_test, _) = mnist.load_data() #去掉标签加载数据x_train = x_train.astype('float32') #数据归一化x_test = x_test.astype('float32')x_train /= 255x_test /= 255x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:]))) #把28×28的数据平展成784大小的向量x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))print x_train.shapeprint x_test.shape#建立自编码器模型encoding_dim = 32  #将784维的数据压缩到32维，压缩率为24.5input_img = Input(shape=(784,))# 增加稀疏性encoded = Dense(encoding_dim, activation='relu', activity_regularizer=regularizers.l1(10e-5))(input_img) #对输入的编码表示（压缩）decoded = Dense(784, activation='sigmoid')(encoded) #输入的有损失重构autoencoder = Model(inputs=input_img, outputs=decoded) #把输入映射到它的重构上#建立单独的编码器模型encoder = Model(inputs=input_img, outputs=encoded)#建立单独的解码器模型encoded_input = Input(shape=(encoding_dim,)) #编码维度是解码器的输入decoded_layer = autoencoder.layers[-1]decoder = Model(inputs=encoded_input, outputs=decoded_layer(encoded_input)) #把编码后的数据映射到最后的输出#激活模型（compile）autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy') #配置模型参数，基于元素的二元交叉熵损失，和Adadelta优化算子#训练模型autoencoder.fit(x_train, x_train,          batch_size=batch_size,          epochs=epochs,          shuffle=True,          validation_data=(x_test, x_test))# 可视化encoded_imgs = encoder.predict(x_test)decoded_imgs = decoder.predict(encoded_imgs)n = 10plt.figure(figsize=(20, 4))for i in range(n):    ax = plt.subplot(2, n, i+1) #display original    plt.imshow(x_test[i].reshape(28, 28))    plt.gray()    ax.get_xaxis().set_visible(False)    ax.get_yaxis().set_visible(False)    ax = plt.subplot(2, n, i+1+n) #display reconstruction    plt.imshow(x_test[i].reshape(28, 28))    plt.gray()    ax.get_xaxis().set_visible(False)    ax.get_yaxis().set_visible(False)plt.show()

注意：我看参考链接的正则化都是：
activity_regularizer=regularizers.activity_l1(10e-5))(input_img)
因为运行后报错，我查阅了keras中文文档里正则化的API,发现应该改为
activity_regularizer=regularizers.l1(10e-5))(input_img)

因为模型加入正则化更不容易过拟合，所以训练的epochs从原来的50,加到了100

实验结果：

loss有点大，再找下原因

多层自编码器

为了使loss减小，可以尝试多层训练，设计栈式自编码器
代码如下：

#!/usr/bin/python# -*- coding:utf8 -*-import kerasimport numpy as npfrom keras.datasets import mnistfrom keras.models import Modelfrom keras.layers import Dense, Inputfrom keras import regularizersimport matplotlib.pyplot as pltbatch_size = 256epochs = 100#数据预处理(x_train, _), (x_test, _) = mnist.load_data() #去掉标签加载数据x_train = x_train.astype('float32') #数据归一化x_test = x_test.astype('float32')x_train /= 255x_test /= 255x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:]))) #把28×28的数据平展成784大小的向量x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))print x_train.shapeprint x_test.shape#建立自编码器模型encoding_dim = 32  #将784维的数据压缩到32维，压缩率为24.5input_img = Input(shape=(784,))# 增加稀疏性encoded = Dense(128, activation='relu')(input_img) #对输入的编码表示（压缩）encoded = Dense(64, activation='relu')(encoded)encoded = Dense(encoding_dim, activation='relu')(encoded)decoded = Dense(64, activation='relu')(encoded)decoded = Dense(128, activation='relu')(decoded)decoded = Dense(784, activation='sigmoid')(decoded) #输入的有损失重构autoencoder = Model(inputs=input_img, outputs=decoded) #把输入映射到它的重构上#建立单独的编码器模型encoder = Model(inputs=input_img, outputs=encoded)#建立单独的解码器模型encoded_input = Input(shape=(encoding_dim,)) #编码维度是解码器的输入num_decoder_layers = 3decoder_layer = encoded_inputfor i in range(-num_decoder_layers, 0):    decoder_layer = autoencoder.layers[i](decoder_layer)decoder = Model(encoded_input, decoder_layer) #这里我参考了stack Overflow上的一个回答https://stackoverflow.com/questions/37758496/python-keras-theano-wrong-dimensions-for-deep-autoencoder#激活模型（compile）autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy') #配置模型参数，基于元素的二元交叉熵损失，和Adadelta优化算子#训练模型autoencoder.fit(x_train, x_train,          batch_size=batch_size,          epochs=epochs,          shuffle=True,          validation_data=(x_test, x_test))# 可视化encoded_imgs = encoder.predict(x_test)decoded_imgs = decoder.predict(encoded_imgs)n = 10plt.figure(figsize=(20, 4))for i in range(n):    ax = plt.subplot(2, n, i+1) #display original    plt.imshow(x_test[i].reshape(28, 28))    plt.gray()    ax.get_xaxis().set_visible(False)    ax.get_yaxis().set_visible(False)    ax = plt.subplot(2, n, i+1+n) #display reconstruction    plt.imshow(x_test[i].reshape(28, 28))    plt.gray()    ax.get_xaxis().set_visible(False)    ax.get_yaxis().set_visible(False)plt.show()

注意，由于构建了深度解码器，在单独构建解码器的时候，不能再取最后一层作为解码器了，因为此时最后一层的输入是128维，而此时encoding dim是32维，所以程序会报错https://stackoverflow.com/questions/37758496/python-keras-theano-wrong-dimensions-for-deep-autoencoder

运行结果