o Si @sdZddlZddlZddlZddlmZddlmZddlm Z ej ej Z ej Z ejej e ej e fZddejeefd ejeeffd d ZGd d d ejZGdddejZddZedkriedSdS)z3MS-STFT discriminator, provided here for reference.N)nn) rearrange) NormConv2drr kernel_sizedilationcCs0|dd|dd|dd|ddfS)Nrr)rrr r C/home/ubuntu/.local/lib/python3.10/site-packages/encodec/msstftd.pyget_2d_paddings0r cseZdZdZddddddddgdddd d d d ifd ededededededededejeefdejdejeefdede de de ffdd Z de j fdd ZZS)!DiscriminatorSTFTa3STFT sub-discriminator. Args: filters (int): Number of filters in convolutions in_channels (int): Number of input channels. Default: 1 out_channels (int): Number of output channels. Default: 1 n_fft (int): Size of FFT for each scale. Default: 1024 hop_length (int): Length of hop between STFT windows for each scale. Default: 256 kernel_size (tuple of int): Inner Conv2d kernel sizes. Default: ``(3, 9)`` stride (tuple of int): Inner Conv2d strides. Default: ``(1, 2)`` dilations (list of int): Inner Conv2d dilation on the time dimension. Default: ``[1, 2, 4]`` win_length (int): Window size for each scale. Default: 1024 normalized (bool): Whether to normalize by magnitude after stft. Default: True norm (str): Normalization method. Default: `'weight_norm'` activation (str): Activation function. Default: `'LeakyReLU'` activation_params (dict): Parameters to provide to the activation function. growth (int): Growth factor for the filters. Default: 1 r) )rr )rr T weight_norm LeakyReLUnegative_slopeg?filters in_channels out_channelsn_fft hop_length win_length max_filters filters_scaler dilationsstride normalizednorm activationactivation_paramsc stt| dks Jt| dksJ||_||_||_||_||_||_| |_ t t j |d i||_ tjj|j|j|jt j|j dddd|_d|j}t |_|jt||j| t| dt||j|}t| D](\}}t||d|j|}|jt||| | |dft| |df| d|}qot|t| d|j|}|jt||| d| dft| d| df| dt||j| d| dft| d| df| d|_dS) Nr F)rrr window_fnr centerpad_modepower)rpaddingr)rrrr(r!r)rr(r!r )super__init__lenrrrrrrr getattrtorchrr" torchaudio transforms Spectrogram hann_windowspec_transform ModuleListconvsappendrr min enumerate conv_post)selfrrrrrrrrrrrr r!r"r# spec_channelsin_chsirout_chs __class__r r r*.sL     zDiscriminatorSTFT.__init__xcCspg}||}tj|j|jgdd}t|d}t|jD]\}}||}||}| |q| |}||fS)Nr)dimzb c w t -> b c t w) r2r-catrealimagrr7r4r"r5r8)r9r@fmapzr<layerr r r forwardVs     zDiscriminatorSTFT.forward)__name__ __module__ __qualname____doc__inttpTupleListboolstrdictr*r-TensorrH __classcell__r r r>r r sB   (r csxeZdZdZddgdgdgdfdedededejed ejed ejef fd d Zd ej de fddZ Z S)MultiScaleSTFTDiscriminatoraMulti-Scale STFT (MS-STFT) discriminator. Args: filters (int): Number of filters in convolutions in_channels (int): Number of input channels. Default: 1 out_channels (int): Number of output channels. Default: 1 n_ffts (Sequence[int]): Size of FFT for each scale hop_lengths (Sequence[int]): Length of hop between STFT windows for each scale win_lengths (Sequence[int]): Window size for each scale **kwargs: additional args for STFTDiscriminator r)ri)rrWrrrn_ffts hop_lengths win_lengthsc sptttkrtksJJtfddttD|_t|j|_dS)Nc s4g|]}tf|||dqS))rrrrr)r ).0r<rrZrkwargsrYrr[r r ss z8MultiScaleSTFTDiscriminator.__init__..)r)r*r+rr3rangediscriminatorsnum_discriminators)r9rrrrYrZr[r^r>r]r r*ns (  z$MultiScaleSTFTDiscriminator.__init__r@returncCs<g}g}|jD]}||\}}||||q||fS)N)rar5)r9r@logitsfmapsdisclogitrEr r r rHzs    z#MultiScaleSTFTDiscriminator.forward) rIrJrKrLrMrNrPr*r-rTDiscriminatorOutputrHrUr r r>r rVcs  rVcCstdd}tddd}tddd}||\}}||\}}t|t|kr=t|kr=t|kr=|jks@JJtdd||DsMJtdd||DsZJtdd||DsgJdS) N )rri]cSsg|]}t|dkqS))r+)r\fmr r r r_sztest..cSs0g|]}|D]}t|jddddgkqqS)Nr rri)listshape)r\rkfr r r r_s0cSsg|] }t|jdkqS)r)r+rm)r\rdr r r r_s)rVr-randnr+rball)rfyy_haty_disc_rfmap_r y_disc_genfmap_genr r r tests   Brw__main__)r)rLtypingrNr.r-reinopsrmodulesrrPrTFeatureMapType LogitsTyperOrhrMr Moduler rVrwrIr r r r s"    (G!