o )jj@sddlZddlZddlZddlZddlmZddlmmZ ddlm Z m Z ddl m Z dZd0ddZdd ZGd d d ejZGd d d e ZGdddejZGdddejjZGdddejZd1ddZd2ddZGdddejjZGdddejjZGdddejZ Gd d!d!ejjZGd"d#d#ejjZGd$d%d%ejjZGd&d'd'ejjZGd(d)d)ejjZ Gd*d+d+ejZ!Gd,d-d-ejZ"Gd.d/d/ejZ#dS)3N)Conv1dConvTranspose1d) Generator皙?csfddfddtfddt|D}t|ddddd f|ddddd f<t|ddd dd f|ddd dd f<|durQd ||<t|S) z Sinusoid position encoding tablecs|tdd|dS)N')nppower)positionZhid_idx)d_hidg/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/modelscope/models/audio/vc/src/vocoder.py cal_anglesz.get_sinusoid_encoding_table..cal_anglecsfddtDS)Ncsg|]}|qSr r ).0Zhid_j)rr r r szKget_sinusoid_encoding_table..get_posi_angle_vec..)ranger )rr rrget_posi_angle_vecsz7get_sinusoid_encoding_table..get_posi_angle_veccsg|]}|qSr r )rZpos_i)rr rrsz/get_sinusoid_encoding_table..Nrrr)r arrayrsincostorchZ FloatTensor)Z n_positionr Z padding_idxZsinusoid_tabler )rr rrget_sinusoid_encoding_tables .. rcCs|dd}|dd\}}t||}||}||}||d|}||} |jg|d|R} td| d||} || } | d} |j g|| |R} | j } | d| | | | jg|dR} | S)Nrr)sizemathgcdviewrZarangeZunfoldZ new_tensorlong contiguousZ new_zerosdeviceZ index_add_to)signalZ frame_stepZouter_dimensionsframesZ frame_lengthZsubframe_lengthZ subframe_stepZsubframes_per_frameZ output_sizeZoutput_subframesZsubframe_signalframeresultZdevice_of_resultr r roverlap_and_add's( r)cs$eZdZfddZddZZS) LastLayerc sbtt|ttj|di||_ttj||ddfi||_tjj||||d|_ dS)Nrrbiasr ) superr*__init__getattrrnn activationpadrconv) self in_channels out_channelsnonlinear_activationnonlinear_activation_paramsr2 kernel_size pad_paramsr, __class__r rr.Ds" zLastLayer.__init__cCs"||}||}||}|SN)r1r2r3r4xr r rforwardNs   zLastLayer.forward__name__ __module__ __qualname__r.r@ __classcell__r r r;rr*Bs  r*cs eZdZdZfddZZS) Conv1d1x1z*1x1 Conv1d with customized initialization.cs tt|j||ddd|ddS)zInitialize 1x1 Conv1d module.rr)r9paddingdilationr,N)r-rFr.)r4r5r6r,r;r rr.Xs  zConv1d1x1.__init__)rBrCrD__doc__r.rEr r r;rrFUsrFcs&eZdZdfdd ZddZZS) LastLinearTcsXtt|tjdd|_t||_t|||d|_ t||_ t|||d|_ dS)N皙?)negative_sloper+) r-rJr.r0 LeakyReLUr1Z BatchNorm1dbn_1rFlinear_1bn_2linear_2)r4Zhidden_channel out_channelr,r;r rr.es   zLastLinear.__init__cCs@||}||}||}||}||}||}|Sr=)r1rNrOrPrQr>r r rr@ms      zLastLinear.forwardTrAr r r;rrJcsrJc*eZdZdZdfdd ZddZZS) Stretch2dzStretch2d module.nearestcs$tt|||_||_||_dS)zInitialize Stretch2d module. Args: x_scale (int): X scaling factor (Time axis in spectrogram). y_scale (int): Y scaling factor (Frequency axis in spectrogram). mode (str): Interpolation mode. N)r-rUr.x_scaley_scalemode)r4rWrXrYr;r rr.zs zStretch2d.__init__cCstj||j|jf|jdS)zCalculate forward propagation. Args: x (Tensor): Input tensor (B, C, F, T). Returns: Tensor: Interpolated tensor (B, C, F * y_scale, T * x_scale), )Z scale_factorrY)FZ interpolaterXrWrYr>r r rr@szStretch2d.forward)rVrBrCrDrIr.r@rEr r r;rrUws rUcs*eZdZ  dfdd ZddZZS) UpsampleLayerrTc s<tt|t|ddd|_tj|||||||d|_dS)NrrV)rYrHr,)r-r\r.rUupsampler0rr3) r4Z in_channelrR upsample_rater9striderGrHr,r;r rr.s  zUpsampleLayer.__init__cCs$||d}||d}|S)Nr)r^ unsqueezer3squeezer>r r rr@szUpsampleLayer.forward)rTrAr r r;rr\s  r\r{Gz?cCs.|jj}|ddkr|jj||dSdS)NZConvr)r<rBfindweightdatanormal_)mmeanZstd classnamer r r init_weightssrkcCst|||dSNr)int)r9rHr r r get_paddingsrnc&eZdZdfdd ZddZZS) ResBlock1rrqTcstt|tt|||d|dt||d|dt|||d|dt||d|dt|||d|dt||d|dg|_tt|||ddt|d|dt|||ddt|d|dt|||ddt|d|dg|_dS)NrrrHrGr,r) r-rpr.r0 ModuleListrrnconvs1convs2r4channelsr9rHr,r;r rr.sv    zResBlock1.__init__cCsLt|j|jD]\}}t|t}||}t|t}||}||}q|Sr=)ziprvrwrZ leaky_relu LRELU_SLOPE)r4r?c1c2xtr r rr@s   zResBlock1.forward)rqrrTrAr r r;rrps8rpcro) ResBlock2rqrrqTc sbtt|tt|||d|dt||d|dt|||d|dt||d|dg|_dS)Nrrrt)r-rr.r0rurrnconvsrxr;r rr.s*   zResBlock2.__init__cCs,|jD]}t|t}||}||}q|Sr=)rrZr{r|)r4r?crr r rr@s   zResBlock2.forward)rqrTrAr r r;rrsrcrT)BasisSignalLayerz Basis Signal@csDtt|tj|d|ddd|_t||j_||_ dS)NrrFr+) r-rr.r0Linearrlayer ParameterreL)r4Zbasis_signal_weightrr;r rr.s zBasisSignalLayer.__init__cCs||}t||jd}|Srl)rr)r)r4resourcer r rr@&s zBasisSignalLayer.forward)rr[r r r;rrs rcs6eZdZdZdddddiffdd Zd d ZZS) CausalConv1dz3CausalConv1d module with customized initialization.rTZ ConstantPad1dvaluercsLtt|ttj||d|fi||_tjj|||||d|_dS)zInitialize CausalConv1d module.rr]N) r-rr.r/rr0r2rr3)r4r5r6r9rHr,r2r:r;r rr.2s  zCausalConv1d.__init__cCs,|||ddddd|dfS)zCalculate forward propagation. Args: x (Tensor): Input tensor (B, in_channels, T). Returns: Tensor: Output tensor (B, out_channels, T). Nr)r3r2rr>r r rr@Es,zCausalConv1d.forwardr[r r r;rr/srcs,eZdZdZ dfdd ZddZZS)CausalConvTranspose1dzr r rr@^s$zCausalConvTranspose1d.forwardrSr[r r r;rrOs  rc s@eZdZdZdddddddid id f fd d Zd dZZS) ResidualStackz+Residual stack module introduced in MelGAN.rq rTrMrLrKZReflectionPad1dFc stt|| sT|dddksJdtjttj|d i|ttj||dd|fi|tjj|||||dttj|d i|tjj||d|d|_n-tjttj|d i|t |||||||dttj|d i|tjj||d|d|_tjj||d|d|_ dS) aInitialize ResidualStack module. Args: kernel_size (int): Kernel size of dilation convolution layer. channels (int): Number of channels of convolution layers. dilation (int): Dilation factor. bias (bool): Whether to add bias parameter in convolution layers. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. pad (str): Padding function module name before dilated convolution layer. pad_params (dict): Hyperparameters for padding function. use_causal_conv (bool): Whether to use causal convolution. rrrz$Not support even number kernel size.r]r+)rHr,r2r:Nr ) r-rr.rr0Z Sequentialr/rstackr skip_layer) r4r9ryrHr,r7r8r2r:Zuse_causal_convr;r rr.ksrzResidualStack.__init__cCs||||S)zCalculate forward propagation. Args: c (Tensor): Input tensor (B, channels, T). Returns: Tensor: Output tensor (B, chennels, T). )rr)r4rr r rr@szResidualStack.forwardr[r r r;rrhsArc sveZdZdgdgdddgdgdgdgdgddf fd d Zd d Zd dZddZddZddZZ S)HiFiGANGeneratorPrq )rsrr1) rrrrTc  srtt|t||_t||_t||ddd| d|_|dkr"tnt } t |_ t t||D]B\} \} } |j |durTt|d| |d| d| | d| d| dnt|d| |d| d| | | d| d| d| d q0t |_tt|j D]$} |d| d}t t||D]\}\} }|j| || || d qqt|dddd| d|_||dS) NrrrqrGr,rFrr_r9r`rGr,rGZoutput_paddingr,r+)r-rr.len num_kernels num_upsamplesrconv_prerprr0ruups enumeraterzappendr\r resblocksr conv_postapply_weight_normreset_parameters)r4input_channelsresblock_kernel_sizesupsample_ratesupsample_initial_channel resblock_typeupsample_kernel_sizesresblock_dilation_sizestransposedconvr,resblockiukchjdr;r rr.s\       zHiFiGANGenerator.__init__cCdd}||dS)z:Remove weight normalization module from all of the layers.cSs<ztd|dtjj|WdStyYdSw)NzWeight norm is removed from .)loggingdebugrr0utilsremove_weight_norm ValueErrorrhr r r_remove_weight_norms  z@HiFiGANGenerator.remove_weight_norm.._remove_weight_normNapply)r4rr r rrsz#HiFiGANGenerator.remove_weight_normcCr)z9Apply weight normalization module from all of the layers.cSsDt|tjjst|tjjr tjj|td|ddSdS)NzWeight norm is applied to r) isinstancerr0rrrZ weight_normrrrr r r_apply_weight_norms z>HiFiGANGenerator.apply_weight_norm.._apply_weight_normNr)r4rr r rrsz"HiFiGANGenerator.apply_weight_normcCr)zReset parameters. This initialization follows official implementation manner. https://github.com/descriptinc/melgan-neurips/blob/master/mel2wav/modules.py cSsFt|tjjst|tjjr!|jjddt d|ddSdS)NrrczReset parameters in r) rrr0rrrerfrgrrrr r r_reset_parameters s z._reset_parametersNr)r4rr r rrsz!HiFiGANGenerator.reset_parameterscCs||}t|jD]=}t|t}|j||}d}t|jD]!}|dur3|j||j||}q ||j||j||7}q ||j}q t|}| |}|Sr=) rrrrZr{r|rrrrr4r?rxsrr r rr@s     zHiFiGANGenerator.forwardcCt|tjstj|tjdt|j}| dd d}| |}t |j D]=}t|t}|j||}d}t |jD]!}|durR|j||j||}q?||j||j||7}q?||j}q)t|}||}|SN)ZdtyperrrrZTensorZtensorfloatr$next parametersr# transposerarrrrZr{r|rrrrrr r r inference'(      zHiFiGANGenerator.inference) rBrCrDr.rrrr@rrEr r r;rrs 8  rc sdeZdZdgdddgddddggdgdgdgd d d d f fd d ZdddZddZZS)ConditionGeneratorirrqrrrrrTFc  stt|t||_t||_t||ddd| d|_td|dd|_|dkr*t nt } t j t gd|_t |_tt||D]B\} \}}|j|durht|d | |d | d||d|d | d nt|d | |d | d|||d |d |d | d qDt |_tt|jD]$} |d | d}tt||D]\}\}}|j| |||| d qqt|d ddd| d|_| rtdddd|_nd|_| rt|_dSd|_dS)Nrrrqrr)rrrFrrrr+rrr)Zdim_inZ style_dimZ max_conv_dim)r-rr.rrrrrspk_fcrprrr0rZrandnspk_inforurrrzrr\rrrrrunet FsmnEncoderV2 extra_layer)r4rrrrrrrrr extra_infor,rrrrrrrr;r rr.@sh        zConditionGenerator.__init__N?cCsf|gd}tj|d|jddd}tj|dd}||j}tj|dd}||||}|dur=||} || 7}||||d} t |j D]=} t | t } |j| | } d} t |jD]!} | durx|j| |j| | } qe| |j| |j| | 7} qe| |j} qOt | } || } |jdur|| d|} | d} | gd} | S)N)rrrrrT)dimZkeepdimr)ZpermutersumrarZsoftmaxrrrrrrZr{r|rrrrrrb)r4inpsZextra_mcabZscorerZspk_inprr?rrrr r rr@s4        zConditionGenerator.forwardcCrrrrr r rrrzConditionGenerator.inference)Nrr)rBrCrDr.r@rrEr r r;rr>s ?#rcs2eZdZdZddgdffdd ZddZZS)FeedForwardNetzA two-feed-forward-layer modulerrcsdttj|||d|dddd|_tj|||d|ddddd|_t||_dS)Nrrr)r9rGF)r9rGr,)r-r.r0rw_1w_2Dropoutdropout)r4Zd_inr Zd_outr9rr;r rr.s zFeedForwardNet.__init__cCs@|dd}t||}||}||}|dd}|S)Nrr)rrZZrelurrr)r4r?outputr r rr@s    zFeedForwardNet.forwardr[r r r;rrsrcs(eZdZdfdd ZdddZZS) MemoryBlockV2rc stt|tt|dd}t|dd}|dkr%||7}||8}|||_|_tj|||dd|dd|_ t ||_ dS)NrrrF)groupsr,) r-rr.rmroundlprpr0rconv_dwrr)r4r filter_sizeshiftrZ left_paddingZ right_paddingr;r rr.szMemoryBlockV2.__init__NcCs|dur ||dd}tj|dd|j|jddfddd}||dddd}||7}| |}|durE||dd}|S)NrrZconstantr)rYrrr) Z masked_fillrarZr2rrrr"rr)r4inputmaskr?rr r rr@s   zMemoryBlockV2.forward)rr=rAr r r;rrsrcs8eZdZ        dfd d Zdd d ZZS)rrr0rrrrc stt|||_||_||_||_||_|_t t s+fddt |jD|_t |_t |_t |||_|jt||||dt d|D]} |jt||||dqMt |_t |D]} |jt|||j| |qetj j||dd|_dS)Ncsg|]}qSr r )r_rr rrsz*FsmnEncoderV2.__init__..)rrFr+)r-rr.rfsmn_num_layersnum_memory_units ffn_inner_dimrrrlistrr0ruadapterffn_lstrprojrrmemory_block_lstrrfc) r4rrZ input_dimrrrZspk_dimrrr;rrr.sJ      zFsmnEncoderV2.__init__NcCst||j|j}||}t|j|jD]$\}}||}|||}t||j|j}|d|dkr9||7}q||}t j |dd}|S)Nrrr) rZrZtrainingrrzrrrrrri)r4rrr?ZffnZ memory_blockcontextZmemoryr r rr@/s   zFsmnEncoderV2.forward)rrrrrrrrr=rAr r r;rrs0rr=)rrc)r)$rrnumpyr rZtorch.nnr0Ztorch.nn.functionalZ functionalrZrrZ Starganv3rr|rr)Moduler*rFrJrUr\rkrnrprrrrrrrrrrr r r rs>     D N {#!