Missing Code?

[Coice]

Hello,

Generator1 requires two parameters for forward (x and c) but in the training step only the mel features are used. (No noise, or other features.)

Is this correct?

    def _train_step(self, batch):
        """Train model one step."""
        # parse batch
        x = []

        x.append(batch['feats'])
        embed = batch['embed'].to(self.device)

        y = batch['audios'].to(self.device)
        x = tuple([x_.to(self.device) for x_ in x])
        y_ = self.model["generator"](*x).to(self.device)

Thank you for your time.

[SunMail-hub]

Hi,
Yes, I forgot to add one line when cleaning up the code, and it should be:

 def _train_step(self, batch):
        """Train model one step."""
        # parse batch
        x = []
        x.append(batch['noise'])
        x.append(batch['feats'])

[Coice]

@SunMail-hub thanks for your response.

The eval code seems to have the correct logic, but turning on F0 features or chroma would cause an error:

        """Evaluate model one step."""
        # parse batch
        x = []

        if self.config['use_noise_input']:
            x.append(batch['noise'])
        if self.config['use_f0']:
            x.append(batch['f0_origins'])
        if self.config['use_chroma']:
            x.append(batch['chromas'])
        x.append(batch['feats'])
        y = batch['audios'].to(self.device)
        x = tuple([x_.to(self.device) for x_ in x])
        embed = batch['embed'].to(self.device)
        y_ = self.model["generator"](*x).to(self.device)

Were you concatenating the extra features (f0, chromas, etc) to the mel features to make it one vector for the c parameter or was there some other modification to the Generator1?

Again thanks for your time.

[SunMail-hub]

Hi @Coice, you could see the settings in config file as follows:

use_f0: false
use_chroma: false
use_noise_input: true

Therefore we use c and noise as model input without extra features.