o ̳i2:@sFddlmZddlmZddlmZddlTddlmZm Z ddl m Z dd Z d d Z Gd d d ejZGdddejZGdddejZGdddejZGdddejZGdddejZGdddejZd(dejdedejfddZdejdejfd d!Zdejdejfd"d#ZGd$d%d%ejZGd&d'd'ejZdS))N) rearrange) activations)*)OptionalTuple)parametrizationscO$tj|i|}tj|ddd|SNweightr)namedim)nnConv1dr weight_norm)argskwargsconvrE/home/ubuntu/.local/lib/python3.10/site-packages/xcodec2/vq/module.pyWNConv1drcOr r )rConvTranspose1drr)rrconvtrrrWNConvTranspose1drrcs0eZdZd dedeffdd ZddZZS) ResidualUnitrr dilationc sdtd|d}tttj|dddt||d||dttj|dddt||dd |_dS) NTalpha_logscale activation) kernel_sizerpaddingr)r% super__init__r Sequential Activation1dr SnakeBetarblock)selfr rpad __class__rrr)"s    zResidualUnit.__init__cCs|||SNr-r.xrrrforward,szResidualUnit.forward)rr__name__ __module__ __qualname__intr)r6 __classcell__rrr0rr!s rcs0eZdZd dedeffdd Zdd ZZS) EncoderBlockrrr r stridec sptfdd|D}tjg|ttjddddtdd|||d|ddR|_dS)Ncsg|] }td|dqS)rrr.0dr rr 2sz)EncoderBlock.__init__..rTr r")r%rAr&r')r.r rA dilationsrunitsr0rGrr)0s zEncoderBlock.__init__cC ||Sr2r3r4rrrr6? zEncoderBlock.forward)rrr>r7rrr0rr=/sr=cs4eZdZd dededeffdd Zd d ZZS) DecoderBlockrrr> input_dim output_dimrAc sjttttj|dddt|d|||d|d|dd|_|j fdd|DdS)NTr r"r)r%rAr&output_paddingcsg|]}t|dqS)rBrCrDrPrrrHPsz)DecoderBlock.__init__..) r(r)rr*r+rr,rr-extend)r.rOrPrArIr0rRrr)Cs  zDecoderBlock.__init__cCrKr2r3r4rrrr6RrLzDecoderBlock.forward)rrNrr>r7rrr0rrMBsrMc s>eZdZ   d dedededeffdd Zd d ZZS) ResLSTMrFT dimension num_layers bidirectionalskipcs6t||_tj||s|n|d|d|d|_dS)NrT) batch_firstrW)r(r)rXrLSTMlstm)r.rUrVrWrXr0rrr)Vs  zResLSTM.__init__cCs4t|d}||\}}|jr||}t|d}|S)z[ Args: x: [B, F, T] Returns: y: [B, F, T] zb f t -> b t fzb t f -> b f t)rr[rX)r.r5y_rrrr6`s  zResLSTM.forward)rFTr8r9r:r;boolr)r6r<rrr0rrTUs rTc s\eZdZdZ ddedededeeffdd Zdd ej d eej d ej fd d Z Z S) ConvNeXtBlockaConvNeXt Block adapted from https://github.com/facebookresearch/ConvNeXt to 1D audio signal. Args: dim (int): Number of input channels. intermediate_dim (int): Dimensionality of the intermediate layer. layer_scale_init_value (float, optional): Initial value for the layer scale. None means no scaling. Defaults to None. adanorm_num_embeddings (int, optional): Number of embeddings for AdaLayerNorm. None means non-conditional LayerNorm. Defaults to None. Nr intermediate_dimlayer_scale_init_valueadanorm_num_embeddingscsttj||dd|d|_|du|_|r t||dd|_ntj|dd|_t |||_ t |_ t |||_ |dkrNtj|t|dd|_dSd|_dS) Nr$r?)r%r&groupsư>epsrT requires_grad)r(r)rrdwconvadanorm AdaLayerNormnorm LayerNormLinearpwconv1GELUactpwconv2 Parametertorchonesgamma)r.r rarbrcr0rrr)}s   zConvNeXtBlock.__init__r5cond_embedding_idreturncCs|}||}|dd}|jr|dusJ|||}n||}||}||}||}|jdur;|j|}|dd}||}|S)Nrr)rj transposerkrmrprrrsrw)r.r5rxresidualrrrr6s          zConvNeXtBlock.forwardr2) r8r9r:__doc__r;floatrr)ruTensorr6r<rrr0rr`qs*r`csLeZdZdZd dededeffdd Zdejd ejd ejfd d Z Z S)rlz Adaptive Layer Normalization module with learnable embeddings per `num_embeddings` classes Args: num_embeddings (int): Number of embeddings. embedding_dim (int): Dimension of the embeddings. renum_embeddings embedding_dimrgcs^t||_||_tj||d|_tj||d|_tjj |jj tjj |jj dS)N)rr) r(r)rgr r Embeddingscaleshiftruinitones_r zeros_)r.rrrgr0rrr)s zAdaLayerNorm.__init__r5rxrycCs<||}||}tjj||jf|jd}|||}|S)Nrf)rrr functional layer_normr rg)r.r5rxrrrrrr6s   zAdaLayerNorm.forward)re) r8r9r:r|r;r}r)rur~r6r<rrr0rrls$ rlc seZdZdZ    ddededeeeefd ed eef fd d Zd e j de j fddZ ddZ e ddededefddZZS) ResBlock1a ResBlock adapted from HiFi-GAN V1 (https://github.com/jik876/hifi-gan) with dilated 1D convolutions, but without upsampling layers. Args: dim (int): Number of input channels. kernel_size (int, optional): Size of the convolutional kernel. Defaults to 3. dilation (tuple[int], optional): Dilation factors for the dilated convolutions. Defaults to (1, 3, 5). lrelu_slope (float, optional): Negative slope of the LeakyReLU activation function. Defaults to 0.1. layer_scale_init_value (float, optional): Initial value for the layer scale. None means no scaling. Defaults to None. r?rr?皙?Nr r%r lrelu_sloperbcs|t||_tttj|||d|d|||ddttj|||d|d|||ddttj|||d|d|||ddg|_tttj|||dd||ddttj|||dd||ddttj|||dd||ddg|_ t |durtj |t |dddnd|durtj |t |dddnd|durtj |t |dddndg|_dS)Nrr)rr&rTrh)r(r)rr ModuleListrr get_paddingconvs1convs2 ParameterListrtrurvrw)r.r r%rrrbr0rrr)sj   #    zResBlock1.__init__r5rycCsrt|j|j|jD]-\}}}tjjj||jd}||}tjjj||jd}||}|dur2||}||}q |S)N)negative_slope) ziprrrwrurr leaky_relur)r.r5c1c2rwxtrrrr6s zResBlock1.forwardcCs4|jD]}t|dq|jD]}t|dqdS)Nr )rrremove_parametrizationsr)r.lrrrremove_weight_norm s  zResBlock1.remove_weight_normrcCst|||dS)Nr)r;)r%rrrrr&szResBlock1.get_padding)r?rrN)r)r8r9r:r|r;rr}rr)rur~r6r staticmethodrr<rrr0rrs* C "rHz>r5clip_valrycCsttj||dS)aU Computes the element-wise logarithm of the input tensor with clipping to avoid near-zero values. Args: x (Tensor): Input tensor. clip_val (float, optional): Minimum value to clip the input tensor. Defaults to 1e-7. Returns: Tensor: Element-wise logarithm of the input tensor with clipping applied. )min)rulogclip)r5rrrrsafe_log+s rcCst|t|Sr2)rusignlog1pabsr5rrrsymlog9srcCst|t|dS)Nr)rurexprrrrrsymexp=src @eZdZ  d dededededef fdd Zd d ZZS) SemanticEncoderr?Tinput_channelscode_dimencode_channelsr%biasc tt|tj|||d|dddd|_ttjddtj|||d|dd|dtjddtj|||d|dd|d|_tj|||d|dddd|_ dSNrrF) in_channels out_channelsr%rAr&rT)inplace)r%rAr&r) r(rr)rr initial_convr*ReLUresidual_blocks final_conv)r.rrrr%rr0rrr)CsJ       zSemanticEncoder.__init__cCs&||}|||}||}|S)u 前向传播方法。 Args: x (Tensor): 输入张量，形状为 (Batch, Input_channels, Length) Returns: Tensor: 编码后的张量，形状为 (Batch, Code_dim, Length) rrrr4rrrr6ws  zSemanticEncoder.forwardr?Tr^rrr0rrBs4rc r) SemanticDecoderr?Troutput_channelsdecode_channelsr%rc rr) r(rr)rrrr*rrr)r.rrrr%rr0rrr)s.     zSemanticDecoder.__init__cCs&||}|||}||}|Sr2r)r.zr5rrrr6s  zSemanticDecoder.forwardrr^rrr0rrs&r)r)torch.nnreinopsrralias_free_torchtypingrrtorch.nn.utilsrrrModulerr=rMrTr`rlrrur~r}rrrrrrrrrs(    7iD