o i$@sjddlZddlmZddlmZddlZGdddej Z Gdddej Z Gdddej Z dd Z dS) Nc2eZdZdZfddZddZd ddZZS) AttDotaCompute attention based on dot product. Args: eprojs (int | None): Dimension of input vectors from encoder. dunits (int | None): Dimension of input vectors for decoder. att_dim (int): Dimension of input vectors for attention. csztt||t|||_t|||_Wdn1s$wY||_||_ ||_ d|_ d|_ d|_ dSN)superr__init__ init_scopeLLinearmlp_encmlp_decdunitseprojsatt_dimh_lengthenc_hpre_compute_enc_h)selfr r r __class__^/home/ubuntu/.local/lib/python3.10/site-packages/espnet/nets/chainer_backend/rnn/attentions.pyrs  zAttDot.__init__cCd|_d|_d|_dSz Reset states.Nrrrrrrrreset z AttDot.reset@c Cst|}|jdur"t||_|jjd|_t|j|jdd|_|dur6t |j j ||j ftjd}n|||j }ttt||d|jj}tj|j|dd}t||}tj|jtt|d|jjdd} | |fS)aCompute AttDot forward layer. Args: enc_hs (chainer.Variable | N-dimensional array): Input variable from encoder. dec_z (chainer.Variable | N-dimensional array): Input variable of decoder. scaling (float): Scaling weight to make attention sharp. Returns: chainer.Variable: Weighted sum over flames. chainer.Variable: Attention weight. N n_batch_axesdtypeaxis)lenrF pad_sequencershapertanhr chainerVariablexpzerosr npfloat32reshape broadcast_to expand_dimsr sumsoftmax) renc_hsdec_zatt_prevscalingbatchuewcrrr__call__%s&  zAttDot.__call__r__name__ __module__ __qualname____doc__rrr? __classcell__rrrrrs   rcr) AttLocaCompute location-based attention. Args: eprojs (int | None): Dimension of input vectors from encoder. dunits (int | None): Dimension of input vectors for decoder. att_dim (int): Dimension of input vectors for attention. aconv_chans (int): Number of channels of output arrays from convolutional layer. aconv_filts (int): Size of filters of convolutional layer. cstt||:t|||_tj||dd|_tj||dd|_tj d|dd|dfd|fd|_ t|d|_ Wdn1sHwY||_ ||_ ||_d|_d|_d|_||_dS)NT)nobiasrrr)ksizepad)rrGrrrr r r mlp_att Convolution2Dloc_convgvecr r rrrr aconv_chans)rr r rrO aconv_filtsrrrrbs"   zAttLoc.__init__cCrrrrrrrrurz AttLoc.resetrc sht|}jdurt|_jjd_jjdd_|dur3t j j |j ft jd}n||j }|durSfdd|D}dd|D}t|}||ddj}ttj|dd dd}j|dd}tt|djj}tjjt|j|dddd }t||} tjjtt| djjdd } | | fS) aCompute AttLoc forward layer. Args: enc_hs (chainer.Variable | N-dimensional array): Input variable from encoders. dec_z (chainer.Variable | N-dimensional array): Input variable of decoder. att_prev (chainer.Variable | None): Attention weight. scaling (float): Scaling weight to make attention sharp. Returns: chainer.Variable: Weighted sum over flames. chainer.Variable: Attention weight. Nrrr r"c0g|]}jj|jdd|jdtjdqSrg?r"r-fullr)r/r0.0hhrrr "z#AttLoc.__call__..cSg|]}t|qSrr+r,rVattrrrrXr$)r&rr'r(rr)rr r+r,r-r.r r/r0r1rMswapaxessqueezerKr2r3r rNr*r5r4) rr6r7r8r9r:att_conv dec_z_tiledr<r=r>rrrr?{sB     zAttLoc.__call__r@rArrrrrGVs  rGcs0eZdZdZfddZddZddZZS)NoAttzCompute non-attention layer. This layer is a dummy attention layer to be compatible with other attention-based models. cs*tt|d|_d|_d|_d|_dSr)rrcrrrrr>rrrrrs  zNoAtt.__init__cCsd|_d|_d|_d|_dSr)rrrr>rrrrrs z NoAtt.resetcsjdurt|_jjd_|dur@fdd|D}dd|D}t|}tjjtt|djjdd_ j |fS)amCompute NoAtt forward layer. Args: enc_hs (chainer.Variable | N-dimensional array): Input variable from encoders. dec_z: Dummy. att_prev (chainer.Variable | None): Attention weight. Returns: chainer.Variable: Sum over flames. chainer.Variable: Attention weight. NrcrQrRrSrUrrrrXrYz"NoAtt.__call__..cSrZrr[r\rrrrXr^rr$) rr'r(rr)rr4r2r3r>)rr6r7r8rrrr?s     zNoAtt.__call__rArrrrrcs  rccCsd|jdkrt|j|j|j}|S|jdkr$t|j|j|j|j|j}|S|jdkr.t}|St d)zReturns an attention layer given the program arguments. Args: args (Namespace): The arguments. Returns: chainer.Chain: The corresponding attention module. dotlocationnoattz9chainer supports only noatt, dot, and location attention.) atyperr r adimrGrOrPrcNotImplementedError)argsr]rrratt_fors   rk)r+chainer.functions functionsr' chainer.linkslinksrnumpyr/ChainrrGrcrkrrrrs  Nq :