o ´‹ j¨ ã@s~ddlZddlmZddlmZddlmZddlmZddl m Z Gdd „d ej ƒZ Gd d „d ej ƒZ Gd d „d ej ƒZdS)éN)ÚVersion)Únn)Ú functional)ÚWNé)Ú LayerNormcs(eZdZdZ‡fdd„Zdd„Z‡ZS)ÚResidualConv1dLayerNormBlocka_Conv1d with Layer Normalization and residual connection as in GlowTTS paper. https://arxiv.org/pdf/1811.00002.pdf :: x |-> conv1d -> layer_norm -> relu -> dropout -> + -> o |---------------> conv1d_1x1 ------------------| Args: in_channels (int): number of input tensor channels. hidden_channels (int): number of inner layer channels. out_channels (int): number of output tensor channels. kernel_size (int): kernel size of conv1d filter. num_layers (int): number of blocks. dropout_p (float): dropout rate for each block. c sÞtƒ ¡||_||_||_||_||_||_|dksJdƒ‚|ddks)Jdƒ‚t  ¡|_ t  ¡|_ t |ƒD]}|j   tj|dkrD|n||||dd¡|j   t|ƒ¡q7t ||d¡|_|jjj ¡|jjj ¡dS)Néz$ [!] number of layers should be > 0.rz& [!] kernel size should be odd number.r)Úpadding)ÚsuperÚ__init__Ú in_channelsÚhidden_channelsÚ out_channelsÚ kernel_sizeÚ num_layersÚ dropout_prÚ ModuleListÚ conv_layersÚ norm_layersÚrangeÚappendÚConv1drÚprojÚweightÚdataÚzero_Úbias)Úselfr rrrrrÚidx©Ú __class__©úO/home/kuhnn/.local/lib/python3.10/site-packages/TTS/tts/layers/glow_tts/glow.pyr s*    ÿÿz%ResidualConv1dLayerNormBlock.__init__cCsh|}t|jƒD]!}|j|||ƒ}|j|||ƒ}tjt |¡|j|jd}q||  |¡}||S)úd Shapes: - x: :math:`[B, C, T]` - x_mask: :math:`[B, 1, T]` )Útraining) rrrrÚFÚdropoutÚrelurr%r)rÚxÚx_maskÚx_resÚir"r"r#Úforward7sz$ResidualConv1dLayerNormBlock.forward)Ú__name__Ú __module__Ú __qualname__Ú__doc__r r-Ú __classcell__r"r"r r#r s rcs4eZdZdZd ‡fdd„ Zd dd„Zd d „Z‡ZS) Ú InvConvNearaúInvertible Convolution with input splitting as in GlowTTS paper. https://arxiv.org/pdf/1811.00002.pdf Args: channels (int): input and output channels. num_splits (int): number of splits, also H and W of conv layer. no_jacobian (bool): enable/disable jacobian computations. Note: Split the input into groups of size self.num_splits and perform 1x1 convolution separately. Cast 1x1 conv operation to 2d by reshaping the input for efficiency. éFc sÄtƒ ¡|ddks J‚||_||_||_d|_ttjƒtdƒkr2t  t  |j|j¡  ¡¡d}ntj   t  |j|j¡  ¡d¡d}t  |¡dkrZd|dd…df|dd…df<t |¡|_dS)Nrrz1.9Úcompleteéÿÿÿÿ)r r ÚchannelsÚ num_splitsÚ no_jacobianÚ weight_invrÚtorchÚ __version__ÚqrÚ FloatTensorÚnormal_ÚlinalgÚdetrÚ Parameterr)rr7r8r9ÚkwargsÚw_initr r"r#r Us  " zInvConvNear.__init__Nc Ksn| ¡\}}}||jdksJ‚|dur$d}tj|f|j|jd|}nt |ddg¡}| |d||j|jd|¡}| ddddd¡  ¡ ||j||j|¡}|rl|j dur[|j } nt  |j   ¡¡j|j jd} d} n|j } |jrud} n t |j ¡||j|} |  |j|jdd¡} t || ¡} |  |d|jd||j|¡} |  ddddd¡  ¡ |||¡|} | | fS) r$rNr )ÚdtypeÚdevicerér4©rE)Úsizer8r;ÚonesrErFÚsumÚviewÚpermuteÚ contiguousr:ÚinverserÚfloatÚtor9Úlogdetr&Úconv2d) rr)r*ÚreverserCÚbÚcÚtÚx_lenrrRÚzr"r"r#r-fs,*  $zInvConvNear.forwardcCs0t |j ¡¡j|jjd}tj|dd|_dS)NrHF)Ú requires_grad) r;rOrrPrQrErrBr:)rr:r"r"r#Ú store_inverse‹szInvConvNear.store_inverse)r4F)NF©r.r/r0r1r r-r[r2r"r"r r#r3Fs  %r3cs:eZdZdZ   d ‡fdd„ Zd dd„Zd d „Z‡ZS) Ú CouplingBlockadGlow Affine Coupling block as in GlowTTS paper. https://arxiv.org/pdf/1811.00002.pdf :: x --> x0 -> conv1d -> wavenet -> conv1d --> t, s -> concat(s*x1 + t, x0) -> o '-> x1 - - - - - - - - - - - - - - - - - - - - - - - - - ^ Args: in_channels (int): number of input tensor channels. hidden_channels (int): number of hidden channels. kernel_size (int): WaveNet filter kernel size. dilation_rate (int): rate to increase dilation by each layer in a decoder block. num_layers (int): number of WaveNet layers. c_in_channels (int): number of conditioning input channels. dropout_p (int): wavenet dropout rate. sigmoid_scale (bool): enable/disable sigmoid scaling for output scale. Note: It does not use the conditional inputs differently from WaveGlow. rFc s¬tƒ ¡||_||_||_||_||_||_||_||_ t j   |d|d¡} t j j j | ¡} | |_t j   ||d¡} | jj ¡| jj ¡| |_t|||||||ƒ|_dS)Nrr )r r r rrÚ dilation_raterÚ c_in_channelsrÚ sigmoid_scaler;rrÚutilsÚparametrizationsÚ weight_normÚstartrrrrÚendrÚwn) rr rrr^rr_rr`rdrer r"r#r §s"   zCouplingBlock.__init__NcKs.|durd}|dd…d|jd…f|dd…|jdd…f}}| |¡|}| |||¡}| |¡}|} |dd…d|jd…dd…f} |dd…|jdd…dd…f} |jret dt | d¡¡} |rv|| t | ¡|} d} n| t | ¡||} t  | |ddg¡} t  | | gd¡}|| fS)z‡ Shapes: - x: :math:`[B, C, T]` - x_mask: :math:`[B, 1, T]` - g: :math:`[B, C, 1]` Nr rgíµ ÷Æ°>) r rdrfrer`r;ÚlogÚsigmoidÚexprKÚcat)rr)r*rTÚgrCÚx_0Úx_1ÚoutÚz_0rWÚsÚz_1rRrYr"r"r#r-És$6   zCouplingBlock.forwardcCs|j ¡dS)N)rfÚremove_weight_norm)rr"r"r#r[èszCouplingBlock.store_inverse)rrF)NFNr\r"r"r r#r]s÷ "r])r;Úpackaging.versionrrÚtorch.nnrr&ÚTTS.tts.layers.generic.wavenetrÚgeneric.normalizationrÚModulerr3r]r"r"r"r#Ús     ;J