o j\!@sFddlZddlmmZGdddejjZGdddejjZdS)Ncs:eZdZdZdddddddiffdd Zd d ZZS) KernelPredictorz7Kernel predictor for the location-variable convolutions@ LeakyReLUnegative_slopeg?c st||_||_||_||_||||} ||} |dd} tjtjj ||ddddt tj| di| |_ tjtj |tjj |||| ddt tj| di| tjj |||| ddt tj| di| tj |tjj |||| ddt tj| di| tjj |||| ddt tj| di| tj |tjj |||| ddt tj| di| tjj |||| ddt tj| di| |_ tjj || || dd|_tjj || || dd|_dS)a9 Args: cond_channels (int): number of channel for the conditioning sequence, conv_in_channels (int): number of channel for the input sequence, conv_out_channels (int): number of channel for the output sequence, conv_layers (int): kpnet_ T)paddingbiasN)super__init__conv_in_channelsconv_out_channelsconv_kernel_size conv_layerstorchnn SequentialConv1dgetattr input_convDropout residual_conv kernel_conv bias_conv)self cond_channelsrrrrkpnet_hidden_channelskpnet_conv_size kpnet_dropoutkpnet_nonlinear_activation!kpnet_nonlinear_activation_paramsl_wl_br  __class__r O/home/kuhnn/.local/lib/python3.10/site-packages/TTS/vocoder/layers/lvc_block.pyrs>     zKernelPredictor.__init__c Csx|j\}}}||}|||}||}||}|||j|j|j |j |}|||j|j |}||fS)z~ Args: c (Tensor): the conditioning sequence (batch, cond_channels, cond_length) Returns: ) shaperrrr contiguousviewrrrr) rcbatch_ cond_lengthkbkernelsr r r r)forwardBs    zKernelPredictor.forward)__name__ __module__ __qualname____doc__rr4 __classcell__r r r'r)rs:rcsBeZdZdZ      d fdd Zd d Zed d ZZS)LVCBlockz"the location-variable convolutionsrrrc st||_||_||_tj|_tjj |||d||d|d|dd|_ t ||d|||||| d|_ t |D]!} d| t|dd} tjj|||| d| d} |j| q>dS)Nr ) kernel_sizestrider output_padding)rrrrrr r!r"rr)r=r dilation)rrcond_hop_lengthrrrr ModuleListconvsConvTranspose1dupsamplerkernel_predictorrangeintrappend) r in_channelsrupsample_ratiorrrAr r!r"ir convr'r r)rYs<    zLVCBlock.__init__c Cs|jd}||\}}t|d}||}t|jD]c}t|d}|j||}t|d}|dd|ddddddddf}|dd|ddddf} |||| d|j }|t |ddd|ddft |dd|dddf}q|S)aLforward propagation of the location-variable convolutions. Args: x (Tensor): the input sequence (batch, in_channels, in_length) c (Tensor): the conditioning sequence (batch, cond_channels, cond_length) Returns: Tensor: the output sequence (batch, in_channels, in_length) rg?N) r*rFF leaky_relurErGrrClocation_variable_convolutionrArsigmoidtanh) rxr-rJr3r rLyr1r2r r r)r4s     (FzLVCBlock.forwardc Cs |j\}}}|j\}}}} } || |ks Jd|d| ||t| dd} t|| | fdd}|d|d| |}||krMt|d|fdd}|d||}|ddddddddd|f}|dd }|d | d}td ||} | |d d } | ||d } | S) uperform location-variable convolution operation on the input sequence (x) using the local convolution kernl. Time: 414 μs ± 309 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each), test on NVIDIA V100. Args: x (Tensor): the input sequence (batch, in_channels, in_length). kernel (Tensor): the local convolution kernel (batch, in_channel, out_channels, kernel_size, kernel_length) bias (Tensor): the bias for the local convolution (batch, out_channels, kernel_length) dilation (int): the dilation of convolution. hop_size (int): the hop_size of the conditioning sequence. Returns: (Tensor): the output sequence after performing local convolution. (batch, out_channels, in_length). z&length of (x, kernel) is not matched, z vs rr constantrrNr;zbildsk,biokl->bolsd) r*rHrNpadunfold transposereinsum unsqueezer+r,) rSkernelr r@hop_sizer.r/ in_length out_channelsr= kernel_lengthr or r r)rPs, & z&LVCBlock.location_variable_convolution)r;rr<rrr) r5r6r7r8rr4 staticmethodrPr9r r r'r)r:Vs/r:)rtorch.nn.functionalr functionalrNModulerr:r r r r)sQ