Cumulative normalization, wrong variance computation

asteroid-team / asteroid

The PyTorch-based audio source separation toolkit for researchers

https://asteroid-team.github.io/

MIT License

2.21k stars 419 forks source link

Closed zmolikova closed 1 year ago

zmolikova commented 1 year ago

Hi, I suspect that the computation of variance in CumLN is missing a division by the length of the sequence. That is:

should be cum_var = cum_pow_sum / cnt - cum_mean.pow(2) as in $var(x) = E[x^2] - E[x]^2$.

mpariente commented 1 year ago

You're fully right, thank you !

Could you please submit a PR ?