o i5+@sddlZddlZddlmZddlmZddlZ ddl Z ddlm Z ddl m Z ddlmZGdddejZGdddejZGd d d ejZGd d d ejZd dZdS)N)cuda) _subsamplex)get_vgg2l_odimc*eZdZdZdfdd ZddZZS)RNNPaRNN with projection layer module. Args: idim (int): Dimension of inputs. elayers (int): Number of encoder layers. cdim (int): Number of rnn units. (resulted in cdim * 2 if bidirectional) hdim (int): Number of projection units. subsample (np.ndarray): List to use sabsample the input array. dropout (float): Dropout rate. typ (str): The RNN type. blstmc stt||ddk}|rd|vrtjntj} n d|vr!tjntj} |r(dnd} |=t j |D].} | dkr>|} n|} |rFd|n|} t |d | | | d| ||t |d | t| |q5Wdn1snwY||_| |_||_||_||_||_dS) Nrblstmbirnnrnnz{}{:d}zbt%d)superr__init__L NStepBiLSTM NStepBiGRU NStepLSTMNStepGRU init_scopesixmovesrangesetattrformatLinearelayers rnn_labelcdim subsampletypbidir)selfidimrrhdimrdropoutr r!r riinputdim_cdim __class__\/home/ubuntu/.local/lib/python3.10/site-packages/espnet/nets/chainer_backend/rnn/encoders.pyrs0     z RNNP.__init__cCst|jjdt|tj|jD]N}d|j vr,||j t|dd|\}}}n||j t|d|\}}t ||j |d\}}|dt|t |}t j|t|dddd}qt jt t |t|dddd}t|ts|g}||fS) aIRNNP forward. Args: xs (chainer.Variable): Batch of padded character ids. (B, Tmax) ilens (chainer.Variable): Batch of length of each input batch. (B,) Returns: xs (chainer.Variable):subsampled vector of xs. chainer.Variable: Subsampled vector of ilens.  input lengths: r Nr btraxis)logginginfor*__name__strrrrrr rrrFvstack split_axisnpcumsumtanh isinstancetuple)r"xsilenslayer_ysr+r+r,__call__9s  " * z RNNP.__call__)rr4 __module__ __qualname____doc__rrC __classcell__r+r+r)r,rs rcr)RNNaRNN Module. Args: idim (int): Dimension of the imput. elayers (int): Number of encoder layers. cdim (int): Number of rnn units. hdim (int): Number of projection units. dropout (float): Dropout rate. typ (str): Rnn type. r c stt||ddk}|rd|vrtjntj}n d|vr!tjntj}|r*d|n|} |||||||_ t | ||_ Wdn1sJwY||_ ||_ dS)Nrrr r )rrIrrrrrrrnbrnnrl_lastr r!) r"r#rrr$r%r r!r r(r)r+r,rjs   z RNN.__init__cCst|jjdt|t|}d|jvr"|dd|\}}}n|d|\}}| t |}t j |t |dddd}t j t t |t |dddd}t|ts^|g}||fS)aRBRNN forward propagation. Args: xs (chainer.Variable): Batch of padded character ids. (B, Tmax) ilens (chainer.Variable): Batch of length of each input batch. (B,) Returns: tuple(chainer.Variable): Tuple of `chainer.Variable` objects. chainer.Variable: `ilens` . r-r Nr/rr0)r2r3r*r4r5rto_cpur rJrKr6r7r8r9r:r;r<r=)r"r>r?rArBr+r+r,rCxs  * z RNN.__call__)r rDr+r+r)r,rI]s rIcr)VGG2LzXVGG motibated cnn layers. Args: in_channel (int): Number of channels. r cstt||4tj|ddddd|_tjdddddd|_tjdddddd|_tjdddddd|_ Wdn1sBwY||_ dS)N@r )stridepad) rrMrrr Convolution2Dconv1_1conv1_2conv2_1conv2_2 in_channel)r"rXr)r+r,rs  zVGG2L.__init__c slt|jjdtttj dj d|j j d|j ddt | t | tjdddt |t |tjddd|jj|j|jjtjddtjd|jj|j|jjtjddtjdtddj dj dj dj dfdd ttDfS) aOVGG2L forward propagation. Args: xs (chainer.Variable): Batch of padded character ids. (B, Tmax) ilens (chainer.Variable): Batch of length of each features. (B,) Returns: chainer.Variable: Subsampled vector of xs. chainer.Variable: Subsampled vector of ilens. r-rr r )rP)dtyperOcs&g|]}|d|ddfqS)Nr+).0r&r?r>r+r, s&z"VGG2L.__call__..)r2r3r*r4r5r6 pad_sequenceswapaxesreshapeshaperXrelurTrUmax_pooling_2drVrWxparrayceilr9float32int32rlenr"r>r?r+r[r,rCs8    *zVGG2L.__call__r rDr+r+r)r,rMs rMcs,eZdZdZ dfdd ZddZZS)EncoderaEncoder network class. Args: etype (str): Type of encoder network. idim (int): Number of dimensions of encoder network. elayers (int): Number of layers of encoder network. eunits (int): Number of lstm units of encoder network. eprojs (int): Number of projection units of encoder network. subsample (np.array): Subsampling number. e.g. 1_2_2_2_1 dropout (float): Dropout rate. r c stt||dd} | dvrtd||drn|ddkrJt t |t t ||d|||||| d|_td| d n t t |tt ||d||||| d|_td| d d |_nA|ddkrt t ||||||| d|_t| d nt t|||||| d|_t| d d|_WddSWddS1swYdS)Nvggp)r grurbgruz:Error: need to specify an appropriate encoder architecturer/)rX)r zUse CNN-VGG + z P for encoderz for encoderz( with every-layer projection for encoderz without projection for encoderr )rrkrlstriprstripr2errorr startswithchainer SequentialrMrrencr3upperrIconv_subsampling_factor) r"etyper#reunitseprojsrr%rXr r)r+r,rs`          "zEncoder.__init__cCs|||\}}||fS)aEEncoder forward. Args: xs (chainer.Variable): Batch of padded character ids. (B, Tmax) ilens (chainer.variable): Batch of length of each features. (B,) Returns: chainer.Variable: Output of the encoder. chainer.Variable: (Subsampled) vector of ilens. )rwrir+r+r,rC$s zEncoder.__call__rjrDr+r+r)r,rks 4rkcCst|j||j|j|j||jS)zReturn the Encoder module. Args: idim (int): Dimension of input array. subsample (numpy.array): Subsample number. egs).1_2_2_2_1 Return chainer.nn.Module: Encoder module. )rkrzrr{r| dropout_rate)argsr#rr+r+r, encoder_for5s r)r2ruchainer.functions functionsr6 chainer.linkslinksrnumpyr9rr&espnet.nets.chainer_backend.nets_utilsrespnet.nets.e2e_asr_commonrChainrrIrMrkrr+r+r+r,s     N=H S