Bug with dac.encode(x, n_quantizers)

[ana-kuznetsova]

Hello!,

I have been trying to use both 16k and 44k version of DAC. I am trying to use the same model with the different number of codebooks. E.g. for 16k I use

# Download a model
model_path = dac.utils.download(model_type="16khz")
model = dac.DAC.load(model_path)

model.to('cuda')

# Load audio signal file
audio, sr = torchaudio.load('input.wav')

audio = audio.unsqueeze(0)
z, codes, latents, _, _ = model.encode(audio.to('cuda'), n_quantizers=12)

The shape of the output codes tensor is torch.Size([1, 12, 292]).

However, if I do

z, codes, latents, _, _ = model.encode(audio.to('cuda'), n_quantizers=2)

The output codes still have the same shape, although I would imagine it should be torch.Size([1, 2, 292]) since only two codebooks are used.

[liangjindeamo-yuer]

you should set model.quantizer.training = False