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=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。