ó ¾÷Xc@@sddlmZddlZddlmZddlmZddlmZddlm Z ddlm Z dd l m Z dd l m Z dd lmZeeeeeed „Zd e fd„ƒYZdefd„ƒYZdefd„ƒYZdefd„ƒYZdS(i(tabsolute_importNi(tbackend(t activations(t initializers(t regularizers(t constraints(tLayer(t InputSpec(t interfacesc C@s²|stj|ƒd}n|s8tj|ƒd}n|sTtj|ƒd}n|dk rùd|kowdknrùtjtj|dd…ddd…fd|fƒƒ}tj||ƒ} tj| |ƒ} tj|| |d|ƒ}ntj|d|fƒ}tj||ƒ}|dk rDtj ||ƒ}ntj ƒd kr“tj|tj d||gƒƒ}|j dd|gƒntj|d||fƒ}|S( s&Apply `y . w + b` for every temporal slice y of x. # Arguments x: input tensor. w: weight matrix. b: optional bias vector. dropout: wether to apply dropout (same dropout mask for every temporal slice of the input). input_dim: integer; optional dimensionality of the input. output_dim: integer; optional dimensionality of the output. timesteps: integer; optional number of timesteps. training: training phase tensor or boolean. # Returns Output tensor. iiggð?Niiÿÿÿÿttrainingt tensorflow( tKtshapetNonet ones_liketreshapetdropouttrepeattin_train_phasetdottbias_addRtstackt set_shape( txtwtbRt input_dimt output_dimt timestepsR tonestdropout_matrixtexpanded_dropout_matrix((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyt_time_distributed_denses((: $t RecurrentcB@s•eZdZeeeedd„Zd„Zd„Zd„Zd d„Z d„Z d d„Z d d „Z d d d d „Z d d „Zd „ZRS(sžAbstract base class for recurrent layers. Do not use in a model -- it's not a valid layer! Use its children classes `LSTM`, `GRU` and `SimpleRNN` instead. All recurrent layers (`LSTM`, `GRU`, `SimpleRNN`) also follow the specifications of this class and accept the keyword arguments listed below. # Example ```python # as the first layer in a Sequential model model = Sequential() model.add(LSTM(32, input_shape=(10, 64))) # now model.output_shape == (None, 32) # note: `None` is the batch dimension. # the following is identical: model = Sequential() model.add(LSTM(32, input_dim=64, input_length=10)) # for subsequent layers, not need to specify the input size: model.add(LSTM(16)) ``` # Arguments weights: list of Numpy arrays to set as initial weights. The list should have 3 elements, of shapes: `[(input_dim, output_dim), (output_dim, output_dim), (output_dim,)]`. return_sequences: Boolean. Whether to return the last output in the output sequence, or the full sequence. go_backwards: Boolean (default False). If True, process the input sequence backwards. stateful: Boolean (default False). If True, the last state for each sample at index i in a batch will be used as initial state for the sample of index i in the following batch. unroll: Boolean (default False). If True, the network will be unrolled, else a symbolic loop will be used. Unrolling can speed-up a RNN, although it tends to be more memory-intensive. Unrolling is only suitable for short sequences. implementation: one of {0, 1, or 2}. If set to 0, the RNN will use an implementation that uses fewer, larger matrix products, thus running faster on CPU but consuming more memory. If set to 1, the RNN will use more matrix products, but smaller ones, thus running slower (may actually be faster on GPU) while consuming less memory. If set to 2 (LSTM/GRU only), the RNN will combine the input gate, the forget gate and the output gate into a single matrix, enabling more time-efficient parallelization on the GPU. Note: RNN dropout must be shared for all gates, resulting in a slightly reduced regularization. input_dim: dimensionality of the input (integer). This argument (or alternatively, the keyword argument `input_shape`) is required when using this layer as the first layer in a model. input_length: Length of input sequences, to be specified when it is constant. This argument is required if you are going to connect `Flatten` then `Dense` layers upstream (without it, the shape of the dense outputs cannot be computed). Note that if the recurrent layer is not the first layer in your model, you would need to specify the input length at the level of the first layer (e.g. via the `input_shape` argument) # Input shapes 3D tensor with shape `(batch_size, timesteps, input_dim)`, (Optional) 2D tensors with shape `(batch_size, output_dim)`. # Output shape - if `return_sequences`: 3D tensor with shape `(batch_size, timesteps, units)`. - else, 2D tensor with shape `(batch_size, units)`. # Masking This layer supports masking for input data with a variable number of timesteps. To introduce masks to your data, use an [Embedding](embeddings.md) layer with the `mask_zero` parameter set to `True`. # Note on using statefulness in RNNs You can set RNN layers to be 'stateful', which means that the states computed for the samples in one batch will be reused as initial states for the samples in the next batch. This assumes a one-to-one mapping between samples in different successive batches. To enable statefulness: - specify `stateful=True` in the layer constructor. - specify a fixed batch size for your model, by passing if sequential model: `batch_input_shape=(...)` to the first layer in your model. else for functional model with 1 or more Input layers: `batch_shape=(...)` to all the first layers in your model. This is the expected shape of your inputs *including the batch size*. It should be a tuple of integers, e.g. `(32, 10, 100)`. - specify `shuffle=False` when calling fit(). To reset the states of your model, call `.reset_states()` on either a specific layer, or on your entire model. # Note on specifying initial states in RNNs You can specify the initial state of RNN layers by calling them with the keyword argument `initial_state`. The value of `initial_state` should be a tensor or list of tensors representing the initial state of the RNN layer. icK@s}tt|ƒj|||_||_||_||_||_t|_ t ddƒ|_ d|_ d|_d|_dS(Ntndimii(tsuperR!t__init__treturn_sequencest go_backwardststatefultunrolltimplementationtTruetsupports_maskingRt input_specR t state_specRtrecurrent_dropout(tselfR%R&R'R(R)tkwargs((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR$¯s        cC@sRt|tƒr|d}n|jr=|d|d|jfS|d|jfSdS(Nii(t isinstancetlistR%tunits(R/t input_shape((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pytcompute_output_shapeÁs   cC@s|jr |SdSdS(N(R%R (R/tinputstmask((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyt compute_maskÉs cC@s t‚dS(N(tNotImplementedError(R/R6tstates((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pytstepÏscC@sgS(N((R/R6R ((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyt get_constantsÒscC@sztj|ƒ}tj|ddƒ}tj|ƒ}tj|d|jgƒ}gtt|jƒƒD] }|^qd}|S(Ntaxisii(ii( R t zeros_liketsumt expand_dimsttileR3trangetlenR:(R/R6t initial_statet_tinitial_states((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pytget_initial_statesÕs (cC@s|S(N((R/R6R ((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pytpreprocess_inputÞscK@sÚ|dk rÁt|dƒr´|j}|j}t|tƒsH|g}n|g||_t|ttfƒsy|g}n|gt|ƒ}tt|ƒj ||}||_|S||ds o       = ,t SimpleRNNcB@s}eZdZejdedddd d d d d d d ddd„ƒZd„Zd d„Z d „Z d d „Z d „Z RS( s• Fully-connected RNN where the output is to be fed back to input. # Arguments units: Positive integer, dimensionality of the output space. activation: Activation function to use (see [activations](../activations.md)). If you don't specify anything, no activation is applied (ie. "linear" activation: `a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. kernel_initializer: Initializer for the `kernel` weights matrix, used for the linear transformation of the inputs. (see [initializers](../initializers.md)). recurrent_initializer: Initializer for the `recurrent_kernel` weights matrix, used for the linear transformation of the recurrent state. (see [initializers](../initializers.md)). bias_initializer: Initializer for the bias vector (see [initializers](../initializers.md)). kernel_regularizer: Regularizer function applied to the `kernel` weights matrix (see [regularizer](../regularizers.md)). recurrent_regularizer: Regularizer function applied to the `recurrent_kernel` weights matrix (see [regularizer](../regularizers.md)). bias_regularizer: Regularizer function applied to the bias vector (see [regularizer](../regularizers.md)). activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). (see [regularizer](../regularizers.md)). kernel_constraint: Constraint function applied to the `kernel` weights matrix (see [constraints](../constraints.md)). recurrent_constraint: Constraint function applied to the `recurrent_kernel` weights matrix (see [constraints](../constraints.md)). bias_constraint: Constraint function applied to the bias vector (see [constraints](../constraints.md)). dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs. recurrent_dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state. # References - [A Theoretically Grounded Application of Dropout in Recurrent Neural Networks](http://arxiv.org/abs/1512.05287) ttanhtglorot_uniformt orthogonalR_gcK@s(tt|ƒj|||_tj|ƒ|_||_tj|ƒ|_ tj|ƒ|_ tj|ƒ|_ t j|ƒ|_ t j|ƒ|_t j| ƒ|_t j| ƒ|_tj| ƒ|_tj| ƒ|_tj| ƒ|_tdtd|ƒƒ|_tdtd|ƒƒ|_dS(Ngð?g(R#RqR$R3Rtgett activationtuse_biasRtkernel_initializertrecurrent_initializertbias_initializerRtkernel_regularizertrecurrent_regularizertbias_regularizertactivity_regularizerRtkernel_constrainttrecurrent_constrainttbias_constrainttmintmaxRR.(R/R3RvRwRxRyRzR{R|R}R~RR€RRR.R0((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR$¢s   c C@svt|tƒr|d}n|jr/|dnd}|d|_td|d|jfƒ|_td||jfƒ|_dg|_ |jr|j ƒn|j |j|jfddd|j d|j d|jƒ|_|j |j|jfdd d|jd|jd|jƒ|_|jr`|j |jfdd d|jd|jd|jƒ|_n d|_t|_dS( NiiR tnametkernelt initializert regularizert constrainttrecurrent_kerneltbias(R1R2R'R RRR,R3R-R:Rgt add_weightRxR{RR…RyR|R€R‰RwRzR}RRŠR*tbuilt(R/R4Rd((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pytbuildÈs8             c C@sh|jdkr|Stj|ƒ}|d}|d}t||j|j|j||j|d|ƒSdS(NiiiR (R)R RRR R…RŠRR3(R/R6R R4RR((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRHês  cC@s6|jdkr|}n{d|jko2dknrWtj||d|jƒ}ntj||jƒ}|jdk r“tj||jƒ}n|d}d|jko·dknrÍ||d9}n|tj||j ƒ}|j dk r|j |ƒ}nd|j|jkr)t |_ n||gfS(Niii( R)RR RR…RŠR RR.R‰RvR*RV(R/R6R:tht prev_outputRM((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR;ús     c @s¥g}ˆjdkrÓdˆjko/dknrÓtj|ƒ}|d}tjtj|dd…ddfdƒƒ‰tjˆdt|ƒfƒ‰‡‡fd†}tj|ˆd|ƒ}|j |ƒn|j tj dƒƒdˆj kodknr‹tjtj|dd…ddfd ƒƒ‰tjˆdˆj fƒ‰‡‡fd†}tj|ˆd|ƒ}|j |ƒn|j tj dƒƒ|S( Niiiÿÿÿÿc@stjˆˆjƒS(N(R R((RR/(s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pytdropped_inputssR gð?c@stjˆˆjƒS(N(R RR.((RR/(s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR's(iÿÿÿÿi(iÿÿÿÿi( R)RR RRRRRAtintRRTtcast_to_floatxR.R3( R/R6R RNR4RRtdp_maskt rec_dp_mask((RR/s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR<s,. .  .  cC@s:i|jd6tj|jƒd6|jd6tj|jƒd6tj|jƒd6tj|jƒd6t j|j ƒd6t j|j ƒd6t j|j ƒd 6t j|j ƒd 6tj|jƒd 6tj|jƒd 6tj|jƒd 6|jd6|jd6}tt|ƒjƒ}tt|jƒƒt|jƒƒƒS(NR3RvRwRxRyRzR{R|R}R~RR€RRR.(R3Rt serializeRvRwRRxRyRzRR{R|R}R~RRR€RRR.R#RqRhRiR2Rj(R/RkRl((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRh1s"    N( RmRnRoRtlegacy_recurrent_supportR*R R$RRHR;R<Rh(((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRqqs*/ "   tGRUcB@s€eZdZejddedddd d d d d d d ddd„ƒZd„Zd d „Z d d „Z d „Z d „Z RS(s Gated Recurrent Unit - Cho et al. 2014. # Arguments units: Positive integer, dimensionality of the output space. activation: Activation function to use (see [activations](../activations.md)). If you don't specify anything, no activation is applied (ie. "linear" activation: `a(x) = x`). recurrent_activation: Activation function to use for the recurrent step (see [activations](../activations.md)). use_bias: Boolean, whether the layer uses a bias vector. kernel_initializer: Initializer for the `kernel` weights matrix, used for the linear transformation of the inputs. (see [initializers](../initializers.md)). recurrent_initializer: Initializer for the `recurrent_kernel` weights matrix, used for the linear transformation of the recurrent state. (see [initializers](../initializers.md)). bias_initializer: Initializer for the bias vector (see [initializers](../initializers.md)). kernel_regularizer: Regularizer function applied to the `kernel` weights matrix (see [regularizer](../regularizers.md)). recurrent_regularizer: Regularizer function applied to the `recurrent_kernel` weights matrix (see [regularizer](../regularizers.md)). bias_regularizer: Regularizer function applied to the bias vector (see [regularizer](../regularizers.md)). activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). (see [regularizer](../regularizers.md)). kernel_constraint: Constraint function applied to the `kernel` weights matrix (see [constraints](../constraints.md)). recurrent_constraint: Constraint function applied to the `recurrent_kernel` weights matrix (see [constraints](../constraints.md)). bias_constraint: Constraint function applied to the bias vector (see [constraints](../constraints.md)). dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs. recurrent_dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state. # References - [On the Properties of Neural Machine Translation: Encoder-Decoder Approaches](https://arxiv.org/abs/1409.1259) - [Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling](http://arxiv.org/abs/1412.3555v1) - [A Theoretically Grounded Application of Dropout in Recurrent Neural Networks](http://arxiv.org/abs/1512.05287) Rrt hard_sigmoidRsRtR_gcK@s:tt|ƒj|||_tj|ƒ|_tj|ƒ|_||_t j|ƒ|_ t j|ƒ|_ t j|ƒ|_ t j|ƒ|_t j| ƒ|_t j| ƒ|_t j| ƒ|_tj| ƒ|_tj| ƒ|_tj|ƒ|_tdtd|ƒƒ|_tdtd|ƒƒ|_dS(Ngð?g(R#R—R$R3RRuRvtrecurrent_activationRwRRxRyRzRR{R|R}R~RRR€RR‚RƒRR.(R/R3RvR™RwRxRyRzR{R|R}R~RR€RRR.R0((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR${s"  c C@sát|tƒr|d}n|jr/|dnd}|d|_td|d|jfƒ|_td||jfƒ|_dg|_ |jr|j ƒn|j |j|jdfddd|j d|j d |jƒ|_|j |j|jdfdd d|jd|jd |jƒ|_|jri|j |jdfdd dd d|jd |jƒ|_n d|_|jdd…d|j…f|_|jdd…d|j…f|_|jdd…|j|jd…f|_|jdd…|j|jd…f|_|jdd…|jdd…f|_|jdd…|jdd…f|_|jr¹|j|j |_|j|j|jd!|_|j|jd|_ nd|_d|_d|_ t!|_"dS( NiiR iR„R…R†R‡RˆR‰RŠtzero(#R1R2R'R RRR,R3R-R:RgR‹RxR{RR…RyR|R€R‰RwR}RRŠtkernel_ztrecurrent_kernel_ztkernel_rtrecurrent_kernel_rtkernel_htrecurrent_kernel_htbias_ztbias_rtbias_hR*RŒ(R/R4Rd((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR£sV            %%,))    c C@sæ|jdkrÞtj|ƒ}|d}|d}t||j|j|j||j|d|ƒ}t||j|j |j||j|d|ƒ}t||j |j |j||j|d|ƒ}tj |||gddƒS|SdS(NiiiR R=( R)R RRR R›R¡RR3RR¢RŸR£t concatenate( R/R6R R4RRtx_ztx_rtx_h((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRHØs  c @s g}ˆjdkrìdˆjko/dknrìtj|ƒ}|d}tjtj|dd…ddfd ƒƒ‰tjˆdt|ƒfƒ‰‡‡fd†}gtdƒD]}tj |ˆd|ƒ^q¸}|j |ƒn/|j gtdƒD]}tj dƒ^qÿƒdˆj ko5dknrÖtjtj|dd…ddfd ƒƒ‰tjˆdˆj fƒ‰‡‡fd†}gtdƒD]}tj |ˆd|ƒ^q¢} |j | ƒn/|j gtdƒD]}tj dƒ^qéƒ|S( Niiiÿÿÿÿc@stjˆˆjƒS(N(R R((RR/(s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRósiR gð?c@stjˆˆjƒS(N(R RR.((RR/(s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRs(iÿÿÿÿi(iÿÿÿÿi(R)RR RRRRRAR‘RBRRTR’R.R3( R/R6R RNR4RRRER“R”((RR/s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR<ës(. ../../cC@sÍ|d}|d}|d}|jdkrÌtj||d|jƒ}|jrktj||jƒ}ntj||d|jdd…dd|j…fƒ}|dd…d|j…f}|dd…|jd|j…f} |dd…d|j…f} |dd…|jd|j…f} |j || ƒ} |j | | ƒ} |dd…d|jd…f}tj| ||d|jdd…d|jd…fƒ}|j ||ƒ}n¼|jdkrF|dd…d|j…f}|dd…|jd|j…f} |dd…d|jd…f}nÀ|jdkrútj||d|j ƒ}tj||d|j ƒ} tj||d|j ƒ}|jrtj||jƒ}tj| |jƒ} tj||jƒ}qn tdƒ‚|j |tj||d|jƒƒ} |j | tj||d|jƒƒ} |j |tj| ||d|jƒƒ}| |d| |}d|j|jkrÀt|_n||gfS(NiiisUnknown `implementation` mode.(R)R RR…RwRRŠR‰R3R™RvR›RRŸR¡R¢R£RPRœRžR RR.R*RV(R/R6R:th_tm1R“R”tmatrix_xt matrix_innerR¥R¦t recurrent_zt recurrent_rtztrR§t recurrent_hthhRŽ((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR; sT    )&&#)&&   cC@sMi|jd6tj|jƒd6tj|jƒd6|jd6tj|jƒd6tj|jƒd6tj|j ƒd6t j|j ƒd6t j|j ƒd 6t j|j ƒd 6t j|jƒd 6tj|jƒd 6tj|jƒd 6tj|jƒd6|jd6|jd6}tt|ƒjƒ}tt|jƒƒt|jƒƒƒS(NR3RvR™RwRxRyRzR{R|R}R~RR€RRR.(R3RR•RvR™RwRRxRyRzRR{R|R}R~RRR€RRR.R#R—RhRiR2Rj(R/RkRl((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRh>s$    N( RmRnRoRR–R*R R$RRHR<R;Rh(((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR—Es,4 5  3tLSTMcB@sƒeZdZejddeddded d d d d d d ddd„ƒZd„Zd d „Z d d „Z d „Z d „Z RS(s, Long-Short Term Memory unit - Hochreiter 1997. For a step-by-step description of the algorithm, see [this tutorial](http://deeplearning.net/tutorial/lstm.html). # Arguments units: Positive integer, dimensionality of the output space. activation: Activation function to use (see [activations](../activations.md)). If you don't specify anything, no activation is applied (ie. "linear" activation: `a(x) = x`). recurrent_activation: Activation function to use for the recurrent step (see [activations](../activations.md)). use_bias: Boolean, whether the layer uses a bias vector. kernel_initializer: Initializer for the `kernel` weights matrix, used for the linear transformation of the inputs. (see [initializers](../initializers.md)). recurrent_initializer: Initializer for the `recurrent_kernel` weights matrix, used for the linear transformation of the recurrent state. (see [initializers](../initializers.md)). bias_initializer: Initializer for the bias vector (see [initializers](../initializers.md)). unit_forget_bias: Boolean. If True, add 1 to the bias of the forget gate at initialization. Setting it to true will also force `bias_initializer="zeros"`. This is recommended in [Jozefowicz et al.](http://www.jmlr.org/proceedings/papers/v37/jozefowicz15.pdf) kernel_regularizer: Regularizer function applied to the `kernel` weights matrix (see [regularizer](../regularizers.md)). recurrent_regularizer: Regularizer function applied to the `recurrent_kernel` weights matrix (see [regularizer](../regularizers.md)). bias_regularizer: Regularizer function applied to the bias vector (see [regularizer](../regularizers.md)). activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). (see [regularizer](../regularizers.md)). kernel_constraint: Constraint function applied to the `kernel` weights matrix (see [constraints](../constraints.md)). recurrent_constraint: Constraint function applied to the `recurrent_kernel` weights matrix (see [constraints](../constraints.md)). bias_constraint: Constraint function applied to the bias vector (see [constraints](../constraints.md)). dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs. recurrent_dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state. # References - [Long short-term memory](http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdf) (original 1997 paper) - [Learning to forget: Continual prediction with LSTM](http://www.mitpressjournals.org/doi/pdf/10.1162/089976600300015015) - [Supervised sequence labeling with recurrent neural networks](http://www.cs.toronto.edu/~graves/preprint.pdf) - [A Theoretically Grounded Application of Dropout in Recurrent Neural Networks](http://arxiv.org/abs/1512.05287) RrR˜RsRtR_gcK@sCtt|ƒj|||_tj|ƒ|_tj|ƒ|_||_t j|ƒ|_ t j|ƒ|_ t j|ƒ|_ ||_ tj| ƒ|_tj| ƒ|_tj| ƒ|_tj| ƒ|_tj| ƒ|_tj|ƒ|_tj|ƒ|_tdtd|ƒƒ|_tdtd|ƒƒ|_dS(Ngð?g(R#R±R$R3RRuRvR™RwRRxRyRztunit_forget_biasRR{R|R}R~RRR€RR‚RƒRR.(R/R3RvR™RwRxRyRzR²R{R|R}R~RR€RRR.R0((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR$s$   c C@sØt|tƒr|d}n|jr/|dnd}|d|_td|d|jfƒ|_td||jfƒtd||jfƒg|_ddg|_ |jr¸|j ƒn|j |j|jdfddd|j d|j d |jƒ|_|j |j|jdfdd d|jd|jd |jƒ|_|jrÖ|j |jdfdd d|jd|jd |jƒ|_|jrßtj|jdfƒ}d ||j|jd+tj|j|ƒqßn d|_|jdd…d|j…f|_|jdd…|j|jd…f|_|jdd…|jd|jd …f|_ |jdd…|jd d…f|_!|jdd…d|j…f|_"|jdd…|j|jd…f|_#|jdd…|jd|jd …f|_$|jdd…|jd d…f|_%|jr§|j|j |_&|j|j|jd!|_'|j|jd|jd !|_(|j|jd |_)n$d|_&d|_'d|_(d|_)t*|_+dS(NiiR iR„R…R†R‡RˆR‰RŠgð?i(,R1R2R'R RRR,R3R-R:RgR‹RxR{RR…RyR|R€R‰RwRzR}RRŠR²RaR_R Rbtkernel_itkernel_ftkernel_ctkernel_otrecurrent_kernel_itrecurrent_kernel_ftrecurrent_kernel_ctrecurrent_kernel_otbias_itbias_ftbias_ctbias_oR*RŒ(R/R4Rdt bias_value((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRºsd             %,0)%,0) !    c C@s|jdkrtj|ƒ}|d}|d}t||j|j|j||j|d|ƒ}t||j|j |j||j|d|ƒ}t||j |j |j||j|d|ƒ}t||j |j |j||j|d|ƒ} tj|||| gddƒS|SdS(NiiiR R=(R)R RRR R³R»RR3R´R¼RµR½R¶R¾R¤( R/R6R R4RRtx_itx_ftx_ctx_o((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRH÷s$  c @s g}ˆjdkrìdˆjko/dknrìtj|ƒ}|d}tjtj|dd…ddfd ƒƒ‰tjˆdt|ƒfƒ‰‡‡fd†}gtdƒD]}tj |ˆd|ƒ^q¸}|j |ƒn/|j gtdƒD]}tj dƒ^qÿƒdˆj ko5dknrÖtjtj|dd…ddfd ƒƒ‰tjˆdˆj fƒ‰‡‡fd†}gtdƒD]}tj |ˆd|ƒ^q¢} |j | ƒn/|j gtdƒD]}tj dƒ^qéƒ|S( Niiiÿÿÿÿc@stjˆˆjƒS(N(R R((RR/(s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRsiR gð?c@stjˆˆjƒS(N(R RR.((RR/(s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR#s(iÿÿÿÿi(iÿÿÿÿi(R)RR RRRRRAR‘RBRRTR’R.R3( R/R6R RNR4RRRER“R”((RR/s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR< s(. ../../cC@s¾|d}|d}|d}|d}|jdkrstj||d|jƒ}|tj||d|jƒ7}|jr–tj||jƒ}n|dd…d|j…f}|dd…|jd|j…f} |dd…d|jd|j…f} |dd…d|jd…f} |j |ƒ} |j | ƒ} | || |j | ƒ}|j | ƒ}n|jdkr|dd…d|j…f}|dd…|jd|j…f}|dd…d|jd|j…f}|dd…d|jd…f}n®|jdkr¹tj||d|j ƒ|j }tj||d|j ƒ|j}tj||d|jƒ|j}tj||d|jƒ|j}n tdƒ‚|j |tj||d|jƒƒ} |j |tj||d|jƒƒ} | || |j |tj||d|jƒƒ}|j |tj||d|jƒƒ}||j |ƒ}d|j|jkr®t|_n|||gfS(NiiiisUnknown `implementation` mode.(R)R RR…R‰RwRRŠR3R™RvR³R»R´R¼RµR½R¶R¾RPR·R¸R¹RºRR.R*RV(R/R6R:R¨tc_tm1R“R”R­tz0tz1tz2tz3R[tftctoRÀRÁRÂRÃRŽ((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR;-sP    ! &*#&*&$$$' $ cC@sWi|jd6tj|jƒd6tj|jƒd6|jd6tj|jƒd6tj|jƒd6tj|j ƒd6|j d6t j|j ƒd 6t j|j ƒd 6t j|jƒd 6t j|jƒd 6tj|jƒd 6tj|jƒd6tj|jƒd6|jd6|jd6}tt|ƒjƒ}tt|jƒƒt|jƒƒƒS(NR3RvR™RwRxRyRzR²R{R|R}R~RR€RRR.(R3RR•RvR™RwRRxRyRzR²RR{R|R}R~RRR€RRR.R#R±RhRiR2Rj(R/RkRl((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyRh]s&     N( RmRnRoRR–R*R R$RRHR<R;Rh(((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyR±Ss.< =  0(t __future__RtnumpyRatRR RRRRtengineRRtlegacyRR R R!RqR—R±(((s5/tmp/pip-build-isqEY4/keras/keras/layers/recurrent.pyts$ -ÿ4Ôÿ