FIXING ERROR WHILE TRAINING VOCODER MODEL: Target loss not found in the keep_avg_target. You might be exiting the training loop before it is computed or set the target_loss in the model config incorrectly.

 I recently ran into this error while training Univnet and MultibandMelgan vocoder.

ValueError:  [!] Target loss not found in the keep_avg_target. You might be exiting the training loop before it is computed or set the target_loss in the model config incorrectly.

This error in UnivNet and MultibandMelGAN arises because the default target_loss is set to 'loss_0', which represents the discriminator loss. However, the discriminator is not activated until steps_to_start_discriminator (defaulting to 200,000), meaning loss_0 is unavailable during the initial training phase. Consequently, the training process cannot find the specified target loss. Setting target_loss='loss_1' resolves this issue by instructing the training to track the generator loss (loss_1), which is active from the beginning of training, thus ensuring a valid target loss is always available, even before the discriminator starts contributing to the learning process. 

This is the code to do the vocoder training for coqui/TTS,

Maker sure you have this line: target_loss='loss_1' 
 

import os

from trainer import Trainer, TrainerArgs

from TTS.utils.audio import AudioProcessor

from TTS.vocoder.configs import UnivnetConfig

from TTS.vocoder.datasets.preprocess import load_wav_data

from TTS.vocoder.models.gan import GAN

output_path = "/content/drive/MyDrive/vocoder_train_run"

if not os.path.exists(output_path):

    os.makedirs(output_path)

#output_path = os.path.dirname(os.path.abspath(__file__))

config = UnivnetConfig(

    batch_size=64,

    eval_batch_size=16,

    num_loader_workers=4,

    num_eval_loader_workers=4,

    run_eval=True,

    test_delay_epochs=-1,

    epochs=1000,

    seq_len=8192,

    pad_short=2000,

    use_noise_augment=True,

    eval_split_size=10,

    print_step=25,

    print_eval=False,

    mixed_precision=False,

    lr_gen=1e-4,

    lr_disc=1e-4,

    data_path=dataset_config.path,

    output_path=output_path,

    target_loss='loss_1' #####

)

# init audio processor

ap = AudioProcessor(**config.audio.to_dict())

# load training samples

eval_samples, train_samples = load_wav_data(config.data_path, config.eval_split_size)

# init model

model = GAN(config, ap)

# init the trainer

trainer = Trainer(

    TrainerArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples

)

 

 

trainer.fit()