o
    
j0                     @   s4   d dl mZmZ ddlmZ eG dd deZdS )    )	dataclassfield   )BaseGANVocoderConfigc                   @   sD  e Zd ZU dZdZeed< dZeed< edd dZ	e
ed	< d
Zeed< edd dZe
ed< dZeed< dZeed< dZeed< dZeed< dZeed< dZeed< dZeed< dZeed< dZeed< dZeed< dZeed< edd dZe
ed< d Zeed!< d"Zeed#< d$Zeed%< d"Zeed&< d'Z eed(< d)Z!eed*< d+S ),FullbandMelganConfiga	  Defines parameters for FullBand MelGAN vocoder.

    Example:

        >>> from TTS.vocoder.configs import FullbandMelganConfig
        >>> config = FullbandMelganConfig()

    Args:
        model (str):
            Model name used for selecting the right model at initialization. Defaults to `fullband_melgan`.
        discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
            'melgan_multiscale_discriminator`.
        discriminator_model_params (dict): The discriminator model parameters. Defaults to
            '{"base_channels": 16, "max_channels": 1024, "downsample_factors": [4, 4, 4, 4]}`
        generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
            considered as a generator too. Defaults to `melgan_generator`.
        batch_size (int):
            Batch size used at training. Larger values use more memory. Defaults to 16.
        seq_len (int):
            Audio segment length used at training. Larger values use more memory. Defaults to 8192.
        pad_short (int):
            Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
        use_noise_augment (bool):
            enable / disable random noise added to the input waveform. The noise is added after computing the
            features. Defaults to True.
        use_cache (bool):
            enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
            not large enough. Defaults to True.
        use_stft_loss (bool):
            enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
        use_subband_stft (bool):
            enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
        use_mse_gan_loss (bool):
            enable / disable using Mean Squeare Error GAN loss. Defaults to True.
        use_hinge_gan_loss (bool):
            enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
            Defaults to False.
        use_feat_match_loss (bool):
            enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
        use_l1_spec_loss (bool):
            enable / disable using L1 spectrogram loss originally used by HifiGAN model. Defaults to False.
        stft_loss_params (dict): STFT loss parameters. Default to
        `{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`
        stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
            model loss. Defaults to 0.5.
        subband_stft_loss_weight (float):
            Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
        mse_G_loss_weight (float):
            MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
        hinge_G_loss_weight (float):
            Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
        feat_match_loss_weight (float):
            Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 108.
        l1_spec_loss_weight (float):
            L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
    fullband_melganmodelmelgan_multiscale_discriminatordiscriminator_modelc                   C   s   ddg ddS )N      )   r   r   )base_channelsmax_channelsdownsample_factors r   r   r   ]/home/kuhnn/.local/lib/python3.10/site-packages/TTS/vocoder/configs/fullband_melgan_config.py<lambda>F   s    zFullbandMelganConfig.<lambda>)default_factorydiscriminator_model_paramsmelgan_generatorgenerator_modelc                   C   s   g dddS )N)   r      r   r   )upsample_factorsnum_res_blocksr   r   r   r   r   r   J   s    generator_model_paramsr   
batch_sizei    seq_leni  	pad_shortTuse_noise_augment	use_cacheuse_stft_lossFuse_subband_stft_lossuse_mse_gan_lossuse_hinge_gan_lossuse_feat_match_lossuse_l1_spec_lossc                   C   s   g dg dg ddS )N)i   i   r   )x      2   )iX  i  r)   )n_fftshop_lengthswin_lengthsr   r   r   r   r   r   ]   s   stft_loss_paramsg      ?stft_loss_weightr   subband_stft_loss_weightg      @mse_G_loss_weighthinge_G_loss_weightl   feat_match_loss_weightg        l1_spec_loss_weightN)"__name__
__module____qualname____doc__r   str__annotations__r
   r   r   dictr   r   r   intr   r   r    boolr!   r"   r#   r$   r%   r&   r'   r.   r/   floatr0   r1   r2   r4   r5   r   r   r   r   r      s>   
 9	r   N)dataclassesr   r   shared_configsr   r   r   r   r   r   <module>   s    