o i$@sdZddlZddlmZmZmZddlZddlmZGdddej j Z Gdddej j Z Gd d d ej j Z Gd d d ej j ZGd ddej j ZGdddej j ZdS)z_Basic Flow modules used in VITS. This code is based on https://github.com/jaywalnut310/vits. N)OptionalTupleUnion)&piecewise_rational_quadratic_transformc @sDeZdZdZdddejdedeejeejejfffddZ d S) FlipFlowzFlip flow module.F)inversexrreturncOs.t|dg}|s||d}||fS|S)a6Calculate forward propagation. Args: x (Tensor): Input tensor (B, channels, T). inverse (bool): Whether to inverse the flow. Returns: Tensor: Flipped tensor (B, channels, T). Tensor: Log-determinant tensor for NLL (B,) if not inverse. r)torchflip new_zerossize)selfrrargskwargslogdetrM/home/ubuntu/.local/lib/python3.10/site-packages/espnet2/gan_tts/vits/flow.pyforwards zFlipFlow.forwardN) __name__ __module__ __qualname____doc__r Tensorboolrrrrrrrrsrc@sNeZdZdZ  d dejdejdededeeje ejejfff d d Z d S) LogFlowzLog flow module.Fh㈵>rx_maskrepsr cKsF|stt|||}t| ddg}||fSt||}|S)aCalculate forward propagation. Args: x (Tensor): Input tensor (B, channels, T). x_mask (Tensor): Mask tensor (B, 1, T). inverse (bool): Whether to inverse the flow. eps (float): Epsilon for log. Returns: Tensor: Output tensor (B, channels, T). Tensor: Log-determinant tensor for NLL (B,) if not inverse. r )r log clamp_minsumexp)rrrrrryrrrrr.s zLogFlow.forwardN)Fr) rrrrr rrfloatrrrrrrrr+src s^eZdZdZdeffdd Z d dejdejded e eje ejejfffd d Z Z S) ElementwiseAffineFlowzElementwise affine flow module.channelsc sLt||_|dtjt|d|dtjt|ddS)zqInitialize ElementwiseAffineFlow module. Args: channels (int): Number of channels. mr logsN)super__init__r(register_parameterr nn Parameterzeros)rr( __class__rrr,Os  zElementwiseAffineFlow.__init__Frrrr cKs`|s |jt|j|}||}t|j|ddg}||fS||jt|j |}|S)aiCalculate forward propagation. Args: x (Tensor): Input tensor (B, channels, T). x_lengths (Tensor): Length tensor (B,). inverse (bool): Whether to inverse the flow. Returns: Tensor: Output tensor (B, channels, T). Tensor: Log-determinant tensor for NLL (B,) if not inverse. r r )r)r r$r*r#)rrrrrr%rrrrr[szElementwiseAffineFlow.forward)F) rrrrintr,r rrrrr __classcell__rrr1rr'Ls r'cs@eZdZdZdedeffdd Zdejdejfdd ZZ S) Transposez+Transpose module for torch.nn.Sequential().dim1dim2cst||_||_dS)zInitialize Transpose module.N)r+r,r6r7)rr6r7r1rrr,ws  zTranspose.__init__rr cCs||j|jS)z Transpose.) transposer6r7)rrrrrr}szTranspose.forward) rrrrr3r,r rrr4rrr1rr5tsr5c sfeZdZdZ  ddededededef fd d Z dd ejd ejde ejdejfddZ Z S)DilatedDepthSeparableConvz$Dilated depth-separable conv module.rr( kernel_sizelayers dropout_raterc sttj|_t|D]X}||}|||d}|jtjtjj||||||dt ddtjj ||ddt ddtj tj||dt ddtjj ||ddt ddtj tj | g7_qdS)a.Initialize DilatedDepthSeparableConv module. Args: channels (int): Number of channels. kernel_size (int): Kernel size. layers (int): Number of layers. dropout_rate (float): Dropout rate. eps (float): Epsilon for layer norm. r )groupsdilationpaddingr T)relementwise_affineN) r+r,r r. ModuleListconvsrange SequentialConv1dr5 LayerNormGELUDropout) rr(r;r<r=rir?r@r1rrr,sP     z"DilatedDepthSeparableConv.__init__Nrrgr cCs8|dur||}|jD] }|||}||}q ||S)aACalculate forward propagation. Args: x (Tensor): Input tensor (B, in_channels, T). x_mask (Tensor): Mask tensor (B, 1, T). g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1). Returns: Tensor: Output tensor (B, channels, T). N)rC)rrrrKfr%rrrrs    z!DilatedDepthSeparableConv.forward)r:r)N) rrrrr3r&r,r rrrr4rrr1rr9s2<r9cseZdZdZ  ddededededed ef fd d Z ddejdejde ejde de eje ejejfff ddZ ZS)ConvFlowzConvolutional flow module. @ in_channelshidden_channelsr;r<bins tail_boundcst|d|_||_||_||_tj|j|d|_ t |||dd|_ tj||j|ddd|_ |j j j|j jjdS)a]Initialize ConvFlow module. Args: in_channels (int): Number of input channels. hidden_channels (int): Number of hidden channels. kernel_size (int): Kernel size. layers (int): Number of layers. bins (int): Number of bins. tail_bound (float): Tail bound value. r r r:)r=N)r+r, half_channelsrQrRrSr r.rF input_convr9dds_convprojweightdatazero_bias)rrPrQr;r<rRrSr1rrr,s.  zConvFlow.__init__NFrrrKrr c Cs||ddd\}}||}|j|||d}|||}|j\}} } ||| d| dddd}t |j } |dd|j f| } |d|j d|j f| } |dd|j df}t || | ||d |j d \}}t||gd|}t||ddg}|s||fS|S) aCalculate forward propagation. Args: x (Tensor): Input tensor (B, channels, T). x_mask (Tensor): Mask tensor (B,). g (Optional[Tensor]): Global conditioning tensor (B, channels, 1). inverse (bool): Whether to inverse the flow. Returns: Tensor: Output tensor (B, channels, T). Tensor: Log-determinant tensor for NLL (B,) if not inverse. r r )rKrrT.Nlinear)rtailsrS)splitrrVrWrXshapereshapepermutemathsqrtrQrRrrSr catr#)rrrrKrxaxbhbctdenom unnorm_widthsunnorm_heightsunnorm_derivatives logdet_absrrrrrs0     zConvFlow.forward)rNrO)NF)rrrrr3r&r,r rrrrrrr4rrr1rrMs<1rM)rrdtypingrrrr espnet2.gan_tts.vits.transformrr.Modulerrr'r5r9rMrrrrs !(T