o j@sZddlZddlmZddlmZejjddZGdddejjZGdd d ejZ dS) N)nn) parametrizecCs\|d}||}t|ddd|ddf}t|dd|dddf}||}|S)Nr)torchtanhsigmoid)input_ainput_b n_channelsn_channels_intin_actt_acts_actactsrQ/home/kuhnn/.local/lib/python3.10/site-packages/TTS/tts/layers/generic/wavenet.pyfused_add_tanh_sigmoid_multiplys   rcs:eZdZdZ   d fdd Zd ddZd d ZZS) WNaPWavenet layers with weight norm and no input conditioning. |-----------------------------------------------------------------------------| | |-> tanh -| | res -|- conv1d(dilation) -> dropout -> + -| * -> conv1d1x1 -> split -|- + -> res g -------------------------------------| |-> sigmoid -| | o --------------------------------------------------------------------------- + --------- o Args: in_channels (int): number of input channels. hidden_channes (int): number of hidden channels. kernel_size (int): filter kernel size for the first conv layer. dilation_rate (int): dilations rate to increase dilation per layer. If it is 2, dilations are 1, 2, 4, 8 for the next 4 layers. num_layers (int): number of wavenet layers. c_in_channels (int): number of channels of conditioning input. dropout_p (float): dropout rate. weight_norm (bool): enable/disable weight norm for convolution layers. rTc st|ddks J|ddksJ||_||_||_||_||_||_||_t j |_ t j |_ t ||_|dkrWt j |d||d} t j jjj| dd|_t|D]d} || } t|| | d} | dkr}t j j|d||| | d} n t j j|d||| | d} t j jjj| dd} |j | | |dkrd|}n|}t j ||d}t j jjj|dd}|j |q[|s|dSdS)Nrweight)name)dilationpadding)super__init__ in_channelshidden_channels kernel_size dilation_rate num_layers c_in_channels dropout_prr ModuleList in_layersres_skip_layersDropoutdropoutConv1dutilsparametrizations weight_norm cond_layerrangeintappendremove_weight_norm)selfrrrrrr r!r*r+irrin_layerres_skip_channelsres_skip_layer __class__rrr%sJ         z WN.__init__Nc Ks&t|}t|jg}|durdn|}|dur||}t|jD]l}|j||}||}|durN|d|j} |dd| | d|jddf} nt|} t || |} |j || } ||jdkr|| ddd|jddf|}|| dd|jdddf}q"|| }q"||S)Ng?rr) r zeros_like IntTensorrr+r,rr#r&rr$) r0xx_maskgkwargsoutputn_channels_tensorr1x_in cond_offsetg_lr res_skip_actsrrrforward^s&   &  $" z WN.forwardcCsL|jdkr t|jd|jD]}t|dq|jD]}t|dqdS)Nrr)r rremove_parametrizationsr+r#r$)r0lrrrr/us   zWN.remove_weight_normrrTNN)__name__ __module__ __qualname____doc__rrCr/ __classcell__rrr5rrs 9rcs2eZdZdZ   d fdd Zd ddZZS) WNBlocksaWavenet blocks. Note: After each block dilation resets to 1 and it increases in each block along the dilation rate. Args: in_channels (int): number of input channels. hidden_channes (int): number of hidden channels. kernel_size (int): filter kernel size for the first conv layer. dilation_rate (int): dilations rate to increase dilation per layer. If it is 2, dilations are 1, 2, 4, 8 for the next 4 layers. num_blocks (int): number of wavenet blocks. num_layers (int): number of wavenet layers. c_in_channels (int): number of channels of conditioning input. dropout_p (float): dropout rate. weight_norm (bool): enable/disable weight norm for convolution layers. rTc sVtt|_t|D]} t| dkr|n|||||||| d} |j| qdS)Nr)rrrrrr r!r*)rrrr" wn_blocksr,rr.) r0rrrr num_blocksrr r!r*idxlayerr5rrrs   zWNBlocks.__init__NcCs |}|jD]}||||}q|S)N)rN)r0r9r:r;orQrrrrCs zWNBlocks.forwardrFrG)rHrIrJrKrrCrLrrr5rrM~srM) rrtorch.nn.utilsrjitscriptrModulerrMrrrrs    n