先说结论：不论是machine translation还是music datasets、speech signal modeling，GRU和LSTM的performance差别不大，但GRU往往比LSTM训练时间短、收敛快。

原文：http://blog.csdn.net/meanme/article/details/48845793

1.概要：

传统的RNN在训练long-term dependencies 的时候会遇到很多困难，最常见的便是vanish gradient problen。期间有很多种解决这个问题的方法被发表。大致可以分为两类：一类是以新的方法改善或者代替传统的SGD方法，如Bengio提出的clip gradient；另一种则是设计更加精密的recurrent unit，如LSTM，GRU。而本文的重点是比较LSTM，GRU的performance。由于在machine translation上这两种unit的performance已经得到验证（效果差别不明显，performance差别不大），本文是在music datasets和speech signal modeling上进行实验的。

2.LSTM与GRU：

1） LSTM：

2）GRU：

3）概括的来说，LSTM和GRU都能通过各种Gate将重要特征保留，保证其在long-term 传播的时候也不会被丢失；还有一个不太好理解，作用就是有利于BP的时候不容易vanishing：

3.实验结果：

实验用了三个unit，传统的tanh，以及LSTM和GRU：

可以发现LSTM和GRU的差别并不大，但是都比tanh要明显好很多，所以在选择LSTM或者GRU的时候还要看具体的task data是什么

不过在收敛时间和需要的epoch上，GRU应该要更胜一筹：

4.代码（keras）：

1）LSTM：

class LSTM(Recurrent):'''Acts as a spatiotemporal projection,turning a sequence of vectors into a single vector.Eats inputs with shape:(nb_samples, max_sample_length (samples shorter than this are padded with zeros at the end), input_dim)and returns outputs with shape:if not return_sequences:(nb_samples, output_dim)if return_sequences:(nb_samples, max_sample_length, output_dim)For a step-by-step description of the algorithm, see:http://deeplearning.net/tutorial/lstm.htmlReferences:Long short-term memory (original 97 paper)http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdfLearning to forget: Continual prediction with LSTMhttp://www.mitpressjournals.org/doi/pdf/10.1162/089976600300015015Supervised sequence labelling with recurrent neural networkshttp://www.cs.toronto.edu/~graves/preprint.pdf'''def __init__(self, input_dim, output_dim=128,init='glorot_uniform', inner_init='orthogonal', forget_bias_init='one',activation='tanh', inner_activation='hard_sigmoid',weights=None, truncate_gradient=-1, return_sequences=False):super(LSTM, self).__init__()self.input_dim = input_dimself.output_dim = output_dimself.truncate_gradient = truncate_gradientself.return_sequences = return_sequencesself.init = initializations.get(init)self.inner_init = initializations.get(inner_init)self.forget_bias_init = initializations.get(forget_bias_init)self.activation = activations.get(activation)self.inner_activation = activations.get(inner_activation)self.input = T.tensor3()self.W_i = self.init((self.input_dim, self.output_dim))self.U_i = self.inner_init((self.output_dim, self.output_dim))self.b_i = shared_zeros((self.output_dim))self.W_f = self.init((self.input_dim, self.output_dim))self.U_f = self.inner_init((self.output_dim, self.output_dim))self.b_f = self.forget_bias_init((self.output_dim))self.W_c = self.init((self.input_dim, self.output_dim))self.U_c = self.inner_init((self.output_dim, self.output_dim))self.b_c = shared_zeros((self.output_dim))self.W_o = self.init((self.input_dim, self.output_dim))self.U_o = self.inner_init((self.output_dim, self.output_dim))self.b_o = shared_zeros((self.output_dim))self.params = [self.W_i, self.U_i, self.b_i,self.W_c, self.U_c, self.b_c,self.W_f, self.U_f, self.b_f,self.W_o, self.U_o, self.b_o,]if weights is not None:self.set_weights(weights)def _step(self,xi_t, xf_t, xo_t, xc_t, mask_tm1,h_tm1, c_tm1,u_i, u_f, u_o, u_c):h_mask_tm1 = mask_tm1 * h_tm1c_mask_tm1 = mask_tm1 * c_tm1i_t = self.inner_activation(xi_t + T.dot(h_mask_tm1, u_i))f_t = self.inner_activation(xf_t + T.dot(h_mask_tm1, u_f))c_t = f_t * c_mask_tm1 + i_t * self.activation(xc_t + T.dot(h_mask_tm1, u_c))o_t = self.inner_activation(xo_t + T.dot(h_mask_tm1, u_o))h_t = o_t * self.activation(c_t)return h_t, c_tdef get_output(self, train=False):X = self.get_input(train)padded_mask = self.get_padded_shuffled_mask(train, X, pad=1)X = X.dimshuffle((1, 0, 2))xi = T.dot(X, self.W_i) + self.b_ixf = T.dot(X, self.W_f) + self.b_fxc = T.dot(X, self.W_c) + self.b_cxo = T.dot(X, self.W_o) + self.b_o[outputs, memories], updates = theano.scan(self._step,sequences=[xi, xf, xo, xc, padded_mask],outputs_info=[T.unbroadcast(alloc_zeros_matrix(X.shape[1], self.output_dim), 1),T.unbroadcast(alloc_zeros_matrix(X.shape[1], self.output_dim), 1)],non_sequences=[self.U_i, self.U_f, self.U_o, self.U_c],truncate_gradient=self.truncate_gradient)if self.return_sequences:return outputs.dimshuffle((1, 0, 2))return outputs[-1]def get_config(self):return {"name": self.__class__.__name__,"input_dim": self.input_dim,"output_dim": self.output_dim,"init": self.init.__name__,"inner_init": self.inner_init.__name__,"forget_bias_init": self.forget_bias_init.__name__,"activation": self.activation.__name__,"inner_activation": self.inner_activation.__name__,"truncate_gradient": self.truncate_gradient,"return_sequences": self.return_sequences}

2）GRU：

class GRU(Recurrent):'''Gated Recurrent Unit - Cho et al. 2014Acts as a spatiotemporal projection,turning a sequence of vectors into a single vector.Eats inputs with shape:(nb_samples, max_sample_length (samples shorter than this are padded with zeros at the end), input_dim)and returns outputs with shape:if not return_sequences:(nb_samples, output_dim)if return_sequences:(nb_samples, max_sample_length, output_dim)References:On the Properties of Neural Machine Translation: Encoder–Decoder Approacheshttp://www.aclweb.org/anthology/W14-4012Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modelinghttp://arxiv.org/pdf/1412.3555v1.pdf'''def __init__(self, input_dim, output_dim=128,init='glorot_uniform', inner_init='orthogonal',activation='sigmoid', inner_activation='hard_sigmoid',weights=None, truncate_gradient=-1, return_sequences=False):super(GRU, self).__init__()self.input_dim = input_dimself.output_dim = output_dimself.truncate_gradient = truncate_gradientself.return_sequences = return_sequencesself.init = initializations.get(init)self.inner_init = initializations.get(inner_init)self.activation = activations.get(activation)self.inner_activation = activations.get(inner_activation)self.input = T.tensor3()self.W_z = self.init((self.input_dim, self.output_dim))self.U_z = self.inner_init((self.output_dim, self.output_dim))self.b_z = shared_zeros((self.output_dim))self.W_r = self.init((self.input_dim, self.output_dim))self.U_r = self.inner_init((self.output_dim, self.output_dim))self.b_r = shared_zeros((self.output_dim))self.W_h = self.init((self.input_dim, self.output_dim))self.U_h = self.inner_init((self.output_dim, self.output_dim))self.b_h = shared_zeros((self.output_dim))self.params = [self.W_z, self.U_z, self.b_z,self.W_r, self.U_r, self.b_r,self.W_h, self.U_h, self.b_h,]if weights is not None:self.set_weights(weights)def _step(self,xz_t, xr_t, xh_t, mask_tm1,h_tm1,u_z, u_r, u_h):h_mask_tm1 = mask_tm1 * h_tm1z = self.inner_activation(xz_t + T.dot(h_mask_tm1, u_z))r = self.inner_activation(xr_t + T.dot(h_mask_tm1, u_r))hh_t = self.activation(xh_t + T.dot(r * h_mask_tm1, u_h))h_t = z * h_mask_tm1 + (1 - z) * hh_treturn h_tdef get_output(self, train=False):X = self.get_input(train)padded_mask = self.get_padded_shuffled_mask(train, X, pad=1)X = X.dimshuffle((1, 0, 2))x_z = T.dot(X, self.W_z) + self.b_zx_r = T.dot(X, self.W_r) + self.b_rx_h = T.dot(X, self.W_h) + self.b_houtputs, updates = theano.scan(self._step,sequences=[x_z, x_r, x_h, padded_mask],outputs_info=T.unbroadcast(alloc_zeros_matrix(X.shape[1], self.output_dim), 1),non_sequences=[self.U_z, self.U_r, self.U_h],truncate_gradient=self.truncate_gradient)if self.return_sequences:return outputs.dimshuffle((1, 0, 2))return outputs[-1]def get_config(self):return {"name": self.__class__.__name__,"input_dim": self.input_dim,"output_dim": self.output_dim,"init": self.init.__name__,"inner_init": self.inner_init.__name__,"activation": self.activation.__name__,"inner_activation": self.inner_activation.__name__,"truncate_gradient": self.truncate_gradient,"return_sequences": self.return_sequences}

顶: 0

踩

LSTM对比GRU：Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling相关推荐

A Critical Review of Recurrent Neural Networks for Sequence Learning-论文（综述）阅读笔记
A Critical Review of Recurrent Neural Networks for Sequence Learning 阅读笔记 //2022.3.31 下午15:00开始论文地址 ...
循环神经网络(RNN, Recurrent Neural Networks)介绍
循环神经网络(RNN, Recurrent Neural Networks)介绍循环神经网络(Recurrent Neural Networks,RNNs)已经在众多自然语言处理(Natural ...
Recurrent Neural Networks 简述
前言 RNN:Recurrent Neural Network 是序列数据处理时匹配度最好的模型,现在对当前几个重要的RNN模型做个简单梳理. RNN RNN,循环神经网络,RNN具有天然的时间深度, ...
第五门课序列模型(Sequence Models) 第一周循环序列模型（Recurrent Neural Networks）
第五门课序列模型(Sequence Models) 第一周循环序列模型(Recurrent Neural Networks) 文章目录第五门课序列模型(Sequence Models) 第一周 ...
[C5W1] Sequence Models - Recurrent Neural Networks
第一周循环序列模型(Recurrent Neural Networks) 为什么选择序列模型?(Why Sequence Models?) 在本课程中你将学会序列模型,它是深度学习中最令人激动的内容 ...
吴恩达deeplearning.ai系列课程笔记+编程作业(13)序列模型(Sequence Models)-第一周循环序列模型（Recurrent Neural Networks）
第五门课序列模型(Sequence Models) 第一周循环序列模型(Recurrent Neural Networks) 文章目录第五门课序列模型(Sequence Models) 第一周 ...
【论文阅读】An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
1.摘要一般而言,序列模型与循环网络(recurrent networks)关系紧密(由于RNN的循环自回归结构能较好地表达出时间序列).而传统的卷积网络(convolutional network ...
An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
TCN:An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling 该论 ...
GRU（Gated Recurrent Neural Network）
当时间步数较大或者时间步较小时,循环神经网络的梯度较容易出现衰减或爆炸.虽然裁剪梯度可以应对梯度爆炸,但无法解决梯度衰减的问题. 门控循环神经网络(gated recurrent neural net ...

LSTM对比GRU：Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

1.概要：

2.LSTM与GRU：

1） LSTM：

2）GRU：

3）概括的来说，LSTM和GRU都能通过各种Gate将重要特征保留，保证其在long-term 传播的时候也不会被丢失；还有一个不太好理解，作用就是有利于BP的时候不容易vanishing：

3.实验结果：

实验用了三个unit，传统的tanh，以及LSTM和GRU：

可以发现LSTM和GRU的差别并不大，但是都比tanh要明显好很多，所以在选择LSTM或者GRU的时候还要看具体的task data是什么

不过在收敛时间和需要的epoch上，GRU应该要更胜一筹：

4.代码（keras）：

1）LSTM：

2）GRU：

LSTM对比GRU：Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling相关推荐

最新文章

热门文章