Non-deterministic behaviour with batch size > 1

#9
by prabhatp251 - opened

Hi, I get deterministic tokenisation when I call encode_code with batch size of 1. However, when I use batch size > 1, I pad shorter waveforms with 0.0. I see that the tokenized waveforms are different for padded waveforms compared to when I call with batch size of 1. I expected the tokens to be different towards the right but I see some differences even in the initial tokens.

HKUST Audio org

Hi, this is normal, no need to worry. As long as the reconstructed audio is correct, the differences in tokens are not an issue. Previous codecs, such as Encodec and DAC, exhibit similar behavior.

Sign up or log in to comment