首页 > 代码库 > 深度学习算法实践15---堆叠去噪自动编码机(SdA)原理及实现

深度学习算法实践15---堆叠去噪自动编码机(SdA)原理及实现

我们讨论了去噪自动编码机(dA),并讨论了Theano框架实现的细节。在本节中,我们将讨论去噪自动编码机(dA)的主要应用,即组成堆叠自动编码机(SdA),我们将以MNIST手写字母识别为例,用堆叠自动编码机(SdA)来解决这一问题。

堆叠自动编码机(SdA)是由一系列去噪自动编码机堆叠而成,每个去噪自动编码机的中间层(即编码层)作为下一层的输入层,这样一层一层堆叠起来,构成一个深层网络,这些网络组成堆叠去噪自动编码机(SdA)的表示部分。这部分通过无监督学习,逐层进行培训,每一层均可以还原加入随机噪音后的输入信号,而此时在每个去噪自动编码机(dA)中间层即编码层的输出信号,可以视为原始输入信号的某种表示,是对原始输入信号的某种简化表示。

当将所有去噪自动编机(dA)堆叠形成的网络训练完成之后,再把最后一层的中间层即编码接入逻辑回归网络,作为其输入层,这样就形成了一个新的多层BP网络,隐藏层之间的权值,就是前面利用去噪自动编码机(dA)逐层训练时所得到的权值矩阵。然后将这个网络视为一个标准的BP网络,利用我们原来的BP网络算法,进行监督学习,最后达到我们希望的状态。

可能读者会有疑问,为什么直接就用多层BP网络呢?这样先逐层训练去噪自动编码机(SdA),然后再组成BP网络,进行监督学习,好像很麻烦呀。其实BP网络诞生之初,就有人基于这个做具有多个隐藏层的深度网络了。但是人们很快就发现,基于误差反向传播的BP网各,利用随机梯度下降算法来调整权值,但是随着层数的加深,离输出层越远的隐藏层,其权值调整量将递减,最后导致这种深度网络学习速度非常慢,直接限制了其的使用,因此在深度学习崛起之前,深层网络基本没有实际成功的应用案例。 从我们的堆叠自动编码机(SdA)来看,我们首先通过逐层非监督学习方式训练独立的去噪自动编码机,可以视为神经网络自动发现问题域的特征的过程,通过自动特征提取,来找到解决问题的最优特征。而去噪自动编码机(SdA)的训练,可以视为已经对多层BP网络进行了初步训练,最后的监督学习是对网络权值的微调优化。这样可以较好的解决深度BP网各学习收敛速度慢的问题,使其具有实用价值。

首先定义堆叠去噪自动编码机(SdA)类,代码如下所示:

[python] view plain copy 在CODE上查看代码片派生到我的代码片 from __future__ import print_function    import os  import sys  import timeit    import numpy    import theano  import theano.tensor as T  from theano.tensor.shared_randomstreams import RandomStreams    from logistic_regression import LogisticRegression  from hidden_layer import HiddenLayer  from denosing_autoencoder import DenosingAutoencoder    class SdA(object):      def __init__(          self,          numpy_rng,          theano_rng=None,          n_ins=784,          hidden_layers_sizes=[500, 500],          n_outs=10,          corruption_levels=[0.1, 0.1]      ):          self.sigmoid_layers = []          self.dA_layers = []          self.params = []          self.n_layers = len(hidden_layers_sizes)            assert self.n_layers > 0            if not theano_rng:              theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))          self.x = T.matrix(‘x‘)  # the data is presented as rasterized images          self.y = T.ivector(‘y‘)  # the labels are presented as 1D vector of          for i in range(self.n_layers):              if i == 0:                  input_size = n_ins              else:                  input_size = hidden_layers_sizes[i - 1]              if i == 0:                  layer_input = self.x              else:                  layer_input = self.sigmoid_layers[-1].output                sigmoid_layer = HiddenLayer(rng=numpy_rng,                                          input=layer_input,                                          n_in=input_size,                                          n_out=hidden_layers_sizes[i],                                          activation=T.nnet.sigmoid)              self.sigmoid_layers.append(sigmoid_layer)              self.params.extend(sigmoid_layer.params)              dA_layer = DenosingAutoencoder(numpy_rng=numpy_rng,                            theano_rng=theano_rng,                            input=layer_input,  www.boyuanyl.cn    远博在线 www.lxinyul.cc/    利信娱乐 www.chuangshi88.cn/创世娱乐 www.yxin7.com      易购娱乐                           n_visible=input_size,                            n_hidden=hidden_layers_sizes[i],                            W=sigmoid_layer.W,                            bhid=sigmoid_layer.b)              self.dA_layers.append(dA_layer)          self.logLayer = LogisticRegression(              input=self.sigmoid_layers[-1].output,              n_in=hidden_layers_sizes[-1],              n_out=n_outs          )          self.params.extend(self.logLayer.params)          self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)          self.errors = self.logLayer.errors(self.y)        def pretraining_functions(self, train_set_x, batch_size):          index = T.lscalar(‘index‘)  # index to a minibatch          corruption_level = T.scalar(‘corruption‘)  # % of corruption to use          learning_rate = T.scalar(‘lr‘)  # learning rate to use          batch_begin = index * batch_size          batch_end = batch_begin + batch_size          pretrain_fns = []          for dA in self.dA_layers:              cost, updates = dA.get_cost_updates(corruption_level,                                                  learning_rate)              fn = theano.function(                  inputs=[                      index,                      theano.In(corruption_level, value=http://www.mamicode.com/0.2), theano.In(learning_rate, value=0.1) ], outputs=cost, updates=updates, givens={ self.x: train_set_x[batch_begin: batch_end] } ) pretrain_fns.append(fn) return pretrain_fns def build_finetune_functions(self, datasets, batch_size, learning_rate): (train_set_x, train_set_y) = datasets[0] (valid_set_x, valid_set_y) = datasets[1] (test_set_x, test_set_y) = datasets[2] n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] n_valid_batches //= batch_size n_test_batches = test_set_x.get_value(borrow=True).shape[0] n_test_batches //= batch_size index = T.lscalar(‘index‘) gparams = T.grad(self.finetune_cost, self.params) updates = [ (param, param - gparam * learning_rate) for param, gparam in zip(self.params, gparams) ] train_fn = theano.function( inputs=[index], outputs=self.finetune_cost, updates=updates, givens={ self.x: train_set_x[ index * batch_size: (index + 1) * batch_size ], self.y: train_set_y[ index * batch_size: (index + 1) * batch_size ] }, name=‘train‘ ) test_score_i = theano.function( [index], self.errors, givens={ self.x: test_set_x[ index * batch_size: (index + 1) * batch_size ], self.y: test_set_y[ index * batch_size: (index + 1) * batch_size ] }, name=‘test‘ ) valid_score_i = theano.function( [index], self.errors, givens={ self.x: valid_set_x[ index * batch_size: (index + 1) * batch_size ], self.y: valid_set_y[ index * batch_size: (index + 1) * batch_size ] }, name=‘valid‘ ) def valid_score(): return [valid_score_i(i) for i in range(n_valid_batches)] def test_score(): www.yyzx66.cn/ 易赢在线 www.hsl85.cn/ 创世娱乐 www.yghrcp88.cn/ 华人彩票 www.ycyc66.cn/ 云彩娱乐 return [test_score_i(i) for i in range(n_test_batches)] return train_fn, valid_score, test_score 在构造函数中,n_ins为输入信号维数,hidden_layer_sizes是一个列表,其中每个元素代表一个隐藏层的神经元数量,可以定义多层,例如在上例中,缺省情况下即为两层,n_outs为输出神经元个数,由于是手写数字识别,因此该值为10,corruption_levels是去噪自动编码机(dA)随机噪音级别,上例中分别为10%的随机噪音。 在构造网络过程中,首先建立BP网络的隐藏层,然后权值和Bias与去噪自动编码机(dA)共享,按照缺省参数,会组成一个输入层有584个神经元,第一隐藏层500个神经元,第二个隐藏层500个神经元,输出层为10个神经元,代码中循环部分具体操作如下所示:

i=0时:

    input_size = 584, layer_input = x即为原始输入信号

    BP隐藏层定义:input=x(原始输入信号)n_in=584(28*28),n_out=hidden_layer_sizes[0]=500,激活函数为Sigmoid函数

    dA定义:input=原始输入信号,n_visible=584, n_hidden=hidden_layer_sizes[0]=500,权值与上面定义的隐藏层共享,Bias与上面定义的隐藏层共享

i=1时:

    input_size=500

    layer_input=上一层输出

    BP隐藏层:input=上一层输出,n_in=500,n_out=hidden_layer_sizes[1]=500,激活函数为Sigmoid函数

    dA定义:input=上一层输出,n_visible=500,n_hidden=hidden_layer_sizes[0]=500,权值与上面定义的隐藏层共享,Bias与上面定义的隐藏层共享

至此循环结束,接着定义最后的逻辑回归层:输入层为上面最后一层的输出,输入层节点数为500,输出层节点数为10。

当创建好网络结构之后,SdA类定义了两阶段的训练方法,pretraining_functions用于逐层训练去噪自动编码机(dA),而build_finetune_functions则用于训练BP网络,由于上面的代码与DenosingAutoencoder和MLP类相类似,这里就不再重复介绍了。

下面定义SdAEngine类,用于完成具体的模型训练工作,代码如下所示:

[python] view plain copy 在CODE上查看代码片派生到我的代码片 from __future__ import print_function    import os  import sys  import timeit    import numpy    import theano  import theano.tensor as T  from theano.tensor.shared_randomstreams import RandomStreams    from mnist_loader import MnistLoader  from mlp import HiddenLayer  from sda import SdA      class SdAEngine(object):      def __init__(self):          print(‘create SdAEngine‘)        def train(finetune_lr=0.1, pretraining_epochs=15,                   pretrain_lr=0.001, training_epochs=1000,                   dataset=‘mnist.pkl.gz‘, batch_size=1):          loader = MnistLoader()          datasets = loader.load_data(dataset)          train_set_x, train_set_y = datasets[0]          valid_set_x, valid_set_y = datasets[1]          test_set_x, test_set_y = datasets[2]          n_train_batches = train_set_x.get_value(borrow=True).shape[0]          n_train_batches //= batch_size          numpy_rng = numpy.random.RandomState(89677)          print(‘... building the model‘)          sda = SdA(              numpy_rng=numpy_rng,              n_ins=28 * 28,              hidden_layers_sizes=[1000, 1000, 1000],              n_outs=10          )          print(‘... getting the pretraining functions‘)          pretraining_fns = sda.pretraining_functions(train_set_x=train_set_x,                                                      batch_size=batch_size)          print(‘... pre-training the model‘)          start_time = timeit.default_timer()          corruption_levels = [.1, .2, .3]          for i in range(sda.n_layers):              for epoch in range(pretraining_epochs):                  c = []                  for batch_index in range(n_train_batches):                      c.append(pretraining_fns[i](index=batch_index,                               corruption=corruption_levels[i],                               lr=pretrain_lr))                  print(‘Pre-training layer %i, epoch %d, cost %f‘ % (i, epoch, numpy.mean(c)))          end_time = timeit.default_timer()          print((‘The pretraining code for file ‘ +                 os.path.split(__file__)[1] +                 ‘ ran for %.2fm‘ % ((end_time - start_time) / 60.)), file=sys.stderr)          print(‘... getting the finetuning functions‘)          train_fn, validate_model, test_model = sda.build_finetune_functions(              datasets=datasets,              batch_size=batch_size,              learning_rate=finetune_lr          )          print(‘... finetunning the model‘)          patience = 10 * n_train_batches  # look as this many examples regardless          patience_increase = 2.  # wait this much longer when a new best is                                  # found          improvement_threshold = 0.995  # a relative improvement of this much is                                         # considered significant          validation_frequency = min(n_train_batches, patience // 2)          best_validation_loss = numpy.inf          test_score = 0.          start_time = timeit.default_timer()          done_looping = False          epoch = 0          while (epoch < training_epochs) and (not done_looping):              epoch = epoch + 1              for minibatch_index in range(n_train_batches):                  minibatch_avg_cost = train_fn(minibatch_index)                  iter = (epoch - 1) * n_train_batches + minibatch_index                  if (iter + 1) % validation_frequency == 0:                      validation_losses = validate_model()                      this_validation_loss = numpy.mean(validation_losses)                      print(‘epoch %i, minibatch %i/%i, validation error %f %%‘ %                            (epoch, minibatch_index + 1, n_train_batches,                             this_validation_loss * 100.))                      if this_validation_loss < best_validation_loss:                          if (                              this_validation_loss < best_validation_loss *                              improvement_threshold                          ):                              patience = max(patience, iter * patience_increase)                          best_validation_loss = this_validation_loss                          best_iter = iter                          test_losses = test_model()                          test_score = numpy.mean(test_losses)                          print((‘     epoch %i, minibatch %i/%i, test error of ‘                                 ‘best model %f %%‘) %                                (epoch, minibatch_index + 1, n_train_batches,                                 test_score * 100.))                  if patience <= iter:                      done_looping = True                      break          end_time = timeit.default_timer()          print(              (                  ‘Optimization complete with best validation score of %f %%, ‘                  ‘on iteration %i, ‘                  ‘with test performance %f %%‘              )              % (best_validation_loss * 100., best_iter + 1, test_score * 100.)          )          print((‘The training code for file ‘ +                 os.path.split(__file__)[1] +                 ‘ ran for %.2fm‘ % ((end_time - start_time) / 60.)), file=sys.stderr)  上面的代码基本上是DenosingAutoencoder和MLP训练算法的合成,没有太多可以介绍的部分。 将上面的代码,结合之间介绍的LogisticRegression、HIddenLayer、MnistLoader等类,就可以构成一个完整的堆叠自动编码机(SdA)了。下面是训练网络的代码:

[python] view plain copy 在CODE上查看代码片派生到我的代码片 from sda_engine import SdAEngine    if __name__ == ‘__main__‘:      engine = SdAEngine()      engine.train()  运行上述代码,在我的Mac笔记本上需要跑一个晚上,可以得到识别错误率为1%左右。 大家可以看到,堆叠去噪自动编码机(SdA)训练速度和识别精度方面,与之前介绍的卷积神经网络(CNN)相比,都会有些差距,这就说明不同的网络,适合不同的任务。图像识别领域,首选是卷积神经网络(CNN),而在图像搜索等领域,堆叠去噪自动编码机(SdA)的应用效果更佳。

深度学习算法实践15---堆叠去噪自动编码机(SdA)原理及实现