o
    
j                     @   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                   @   s  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< 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/< d0Z&eed1< ed2d dZ'e
ed3< d4Z(eed5< ed6d dZ)e
ed7< d4Z*eed8< ed9d dZ+e
ed:< dZ,eed;< d<S )=ParallelWaveganConfiga  Defines parameters for ParallelWavegan vocoder.

    Args:
        model (str):
            Model name used for selecting the right configuration at initialization. Defaults to `gan`.
        discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
            'parallel_wavegan_discriminator`.
        discriminator_model_params (dict): The discriminator model kwargs. Defaults to
            '{"num_layers": 10}`
        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 `parallel_wavegan_generator`.
        generator_model_param (dict):
            The generator model kwargs. Defaults to `{"upsample_factors": [4, 4, 4, 4], "stacks": 3, "num_res_blocks": 30}`.
        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.
        steps_to_start_discriminator (int):
            Number of steps required to start training the discriminator. Defaults to 0.
        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 0.
        l1_spec_loss_weight (float):
            L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
        lr_gen (float):
            Generator model initial learning rate. Defaults to 0.0002.
        lr_disc (float):
            Discriminator model initial learning rate. Defaults to 0.0002.
        optimizer (torch.optim.Optimizer):
            Optimizer used for the training. Defaults to `AdamW`.
        optimizer_params (dict):
            Optimizer kwargs. Defaults to `{"betas": [0.8, 0.99], "weight_decay": 0.0}`
        lr_scheduler_gen (torch.optim.Scheduler):
            Learning rate scheduler for the generator. Defaults to `ExponentialLR`.
        lr_scheduler_gen_params (dict):
            Parameters for the generator learning rate scheduler. Defaults to `{"gamma": 0.5, "step_size": 200000, "last_epoch": -1}`.
        lr_scheduler_disc (torch.optim.Scheduler):
            Learning rate scheduler for the discriminator. Defaults to `ExponentialLR`.
        lr_scheduler_dict_params (dict):
            Parameters for the discriminator learning rate scheduler. Defaults to `{"gamma": 0.5, "step_size": 200000, "last_epoch": -1}`.
    parallel_waveganmodelparallel_wavegan_discriminatordiscriminator_modelc                   C   s   ddiS )N
num_layers
    r   r   r   ^/home/kuhnn/.local/lib/python3.10/site-packages/TTS/vocoder/configs/parallel_wavegan_config.py<lambda>T   s    zParallelWaveganConfig.<lambda>)default_factorydiscriminator_model_paramsparallel_wavegan_generatorgenerator_modelc                   C   s   g ddddS )N)   r   r   r         )upsample_factorsstacksnum_res_blocksr   r   r   r   r   r   W   s    generator_model_params   
batch_sizei d  seq_leni  	pad_shortFuse_noise_augmentT	use_cache@ steps_to_start_discriminatorloss_1target_lossuse_stft_loss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   i   )x      2   )iX  i  r,   )n_fftshop_lengthswin_lengthsr   r   r   r   r   r   l   s   stft_loss_params      ?stft_loss_weightr   subband_stft_loss_weightg      @mse_G_loss_weighthinge_G_loss_weightfeat_match_loss_weightl1_spec_loss_weightg-C6*?lr_genlr_discAdamW	optimizerc                   C   s   ddgddS )Ng?gGz?g        )betasweight_decayr   r   r   r   r   r      s    optimizer_paramsStepLRlr_scheduler_genc                   C      ddddS Nr2   r!   )gamma	step_size
last_epochr   r   r   r   r   r          lr_scheduler_gen_paramslr_scheduler_discc                   C   rB   rC   r   r   r   r   r   r      rH   lr_scheduler_disc_paramsscheduler_after_epochN)-__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*   r1   r3   floatr4   r5   r6   r7   r8   r9   r:   r<   r?   rA   rI   rJ   rK   rL   r   r   r   r   r      sT   
 H	r   N)dataclassesr   r   shared_configsr   r   r   r   r   r   <module>   s    