o j!@sLddlmZmZddlmZmZeGdddeZeGdddeZdS)) dataclassfield)BaseAudioConfigBaseTrainingConfigc@seZdZUdZeedZeed<dZe ed<dZ e ed<dZ e ed <d Ze ed <d Ze ed <dZe ed<dZe ed<dZe ed<dZe ed<dZeed<dZe ed<edddZeed<d S)BaseVocoderConfiga"Shared parameters among all the vocoder models. Args: audio (BaseAudioConfig): Audio processor config instance. Defaultsto `BaseAudioConfig()`. use_noise_augment (bool): Augment the input audio with random noise. Defaults to False/ eval_split_size (int): Number of instances used for evaluation. Defaults to 10. data_path (str): Root path of the training data. All the audio files found recursively from this root path are used for training. Defaults to `""`. feature_path (str): Root path to the precomputed feature files. Defaults to None. seq_len (int): Length of the waveform segments used for training. Defaults to 1000. pad_short (int): Extra padding for the waveforms shorter than `seq_len`. Defaults to 0. conv_path (int): Extra padding for the feature frames against convolution of the edge frames. Defaults to MISSING. Defaults to 0. use_cache (bool): enable / disable in memory caching of the computed features. If the RAM is not enough, if may cause OOM. Defaults to False. epochs (int): Number of training epochs to. Defaults to 10000. wd (float): Weight decay. 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}` default_factoryaudioFuse_noise_augment eval_split_size data_pathN feature_pathiseq_lenr pad_shortconv_pad use_cachei'epochswdAdamW optimizercCsddgddS)Ng?gGz?r)betas weight_decayrrrU/home/kuhnn/.local/lib/python3.10/site-packages/TTS/vocoder/configs/shared_configs.py8szBaseVocoderConfig.optimizer_params)__name__ __module__ __qualname____doc__rrr __annotations__r boolr intrstrrrrrrrrfloatrrdictrrrrrs !           rc@seZdZUdZdZeed<dZeed<dZ eed<dZ eed<dZ eed<dZ eed <dZ eed <d Zeed <d Zeed <dZeed<d Zeed<dZeed<dZeed<edddZeed<edddZeed<dZeed<edddZeed<dZeed <dZeed!<d"Zeed#<ed$ddZeed%<d"Zeed&<ed'ddZ eed(<dZ!eed)<d*Z"eed+<d Z#d*Z$eed,<d-S).BaseGANVocoderConfigaBase config class used among all the GAN based vocoders. Args: use_stft_loss (bool): enable / disable the use of STFT loss. Defaults to True. use_subband_stft_loss (bool): enable / disable the use of Subband STFT loss. Defaults to True. use_mse_gan_loss (bool): enable / disable the use of Mean Squared Error based GAN loss. Defaults to True. use_hinge_gan_loss (bool): enable / disable the use of Hinge GAN loss. Defaults to True. use_feat_match_loss (bool): enable / disable feature matching loss. Defaults to True. use_l1_spec_loss (bool): enable / disable L1 spectrogram loss. Defaults to True. stft_loss_weight (float): Loss weight that multiplies the computed loss value. Defaults to 0. subband_stft_loss_weight (float): Loss weight that multiplies the computed loss value. Defaults to 0. mse_G_loss_weight (float): Loss weight that multiplies the computed loss value. Defaults to 1. hinge_G_loss_weight (float): Loss weight that multiplies the computed loss value. Defaults to 0. feat_match_loss_weight (float): Loss weight that multiplies the computed loss value. Defaults to 100. l1_spec_loss_weight (float): Loss weight that multiplies the computed loss value. Defaults to 45. stft_loss_params (dict): Parameters for the STFT loss. Defaults to `{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`. l1_spec_loss_params (dict): Parameters for the L1 spectrogram loss. Defaults to `{ "use_mel": True, "sample_rate": 22050, "n_fft": 1024, "hop_length": 256, "win_length": 1024, "n_mels": 80, "mel_fmin": 0.0, "mel_fmax": None, }` target_loss (str): Target loss name that defines the quality of the model. Defaults to `G_avg_loss`. grad_clip (list): A list of gradient clipping theresholds for each optimizer. Any value less than 0 disables clipping. Defaults to [5, 5]. lr_gen (float): Generator model initial learning rate. Defaults to 0.0002. lr_disc (float): Discriminator model initial learning rate. Defaults to 0.0002. 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": 0.999, "last_epoch": -1}`. lr_scheduler_disc (torch.optim.Scheduler): Learning rate scheduler for the discriminator. Defaults to `ExponentialLR`. lr_scheduler_disc_params (dict): Parameters for the discriminator learning rate scheduler. Defaults to `{"gamma": 0.999, "last_epoch": -1}`. scheduler_after_epoch (bool): Whether to update the learning rate schedulers after each epoch. Defaults to True. use_pqmf (bool): enable / disable PQMF for subband approximation at training. Defaults to False. steps_to_start_discriminator (int): Number of steps required to start training the discriminator. Defaults to 0. diff_samples_for_G_and_D (bool): enable / disable use of different training samples for the generator and the discriminator iterations. Enabling it results in slower iterations but faster convergance in some cases. Defaults to False. ganmodelT use_stft_lossuse_subband_stft_lossuse_mse_gan_lossuse_hinge_gan_lossuse_feat_match_lossuse_l1_spec_lossrstft_loss_weightsubband_stft_loss_weightmse_G_loss_weighthinge_G_loss_weightdfeat_match_loss_weight-l1_spec_loss_weightcCsgdgdgddS)N)ii)x2)iXir=)n_ffts hop_lengths win_lengthsrrrrrrszBaseGANVocoderConfig.rstft_loss_paramsc CsdddddddddS)NTi"Vr;Pr)use_mel sample_raten_fft hop_length win_lengthn_melsmel_fminmel_fmaxrrrrrrsl1_spec_loss_paramsloss_0 target_losscCsddgS)Nrrrrrrs grad_clipg-C6*?lr_genlr_disc ExponentialLRlr_scheduler_gencC dddSNg+?)gamma last_epochrrrrrr lr_scheduler_gen_paramslr_scheduler_disccCrVrWrrrrrrr[lr_scheduler_disc_paramsscheduler_after_epochFuse_pqmfdiff_samples_for_G_and_DN)%rr r!r"r+r&r#r,r$r-r.r/r0r1r2r'r3r5r6r8r:rrBr(rMrOrQrRrSrUr\r]r^r_r`steps_to_start_discriminatorrarrrrr);sB  D                   r)N) dataclassesrr TTS.configrrrr)rrrrs 4