Harmonic component not getting trained while training on a higher sample rate!

[PratikStar]

I am training the vanilla DDSP model (ae.gin), on my custom guitar dataset that has a sample rate= 44.1kHz.

One thing that I have learned is for this high sr; I will have to modify the SpectralLoss to have a higher fft_sizes! But during the training, I noticed that my audio clip is being generated from the filtered noise component and the harmonic component is not getting trained!

I discovered this because I checked the spectrograms of the output signals from harmonic and the noise components! I have trained for about 50k steps, and the loss has been steady for a while now (6.x) I don't understand why this happens and what I should be doing to fix this.

The commands I used: https://github.com/magenta/ddsp/issues/20#issuecomment-1325880712

I am attaching the screenshots of harmonic and noise components here!

(Resynthesized) Harmonic Component

(Resynthesized) Noise Component

(Resynthesized) Harmonic + Noise Component