Skip to content

Conversation

@naxingyu
Copy link
Contributor

@naxingyu naxingyu commented Feb 4, 2020

triggered by [src] changes #3898, #3900, #3901, #3903, #3742, #3904, #3905
@danpovey please merge without squash.

luitjens and others added 8 commits February 1, 2020 11:15
…3898) * removing note in ivector comments and explaining magic constant * Fix some bugs in the online batched cuda ivector code. They are as follows: 1) Striding issue when initializing the gamma stash 2) Rewrote splice kernel. It has small corner case bugs leading to small differences on the boundaries of some chunks. 3) Zero out posteriors for invalid rows. This is necessary because we work with a batch of matrices where each matrix may have different numbers of valid frames. If we don't zero these out X and gamma can have small differences for some gauss points because those invalid posterior rows will add to statistics. Zeroing them out ensures that they don't contribute to the statistics. This now matches to the last siginficant digits across the ivector for multiple datasets: For example: ../ivectorbin/ivector-compute-dot-products trials 'ark:../ivectorbin/ivector-normalize-length --scaleup=false ark:ivector.ark ark:- |' 'ark:../ivectorbin/ivector-normalize-length --scaleup=false ark:ivector-batched.ark ark:- |' scores LOG (ivector-normalize-length[5.5.1697~1-22ac]:main():ivector-normalize-length.cc:90) Processed 80 iVectors. LOG (ivector-normalize-length[5.5.1697~1-22ac]:main():ivector-normalize-length.cc:94) Average ratio of iVector to expected length was 12.214, standard deviation was 7.93438 ../ivectorbin/ivector-normalize-length --scaleup=false ark:ivector-batched.ark ark:- LOG (ivector-normalize-length[5.5.1697~1-22ac]:main():ivector-normalize-length.cc:90) Processed 80 iVectors. LOG (ivector-normalize-length[5.5.1697~1-22ac]:main():ivector-normalize-length.cc:94) Average ratio of iVector to expected length was 12.214, standard deviation was 7.93437 LOG (ivector-compute-dot-products[5.5.1697~1-22ac]:main():ivector-compute-dot-products.cc:108) Mean dot-product was 1, standard deviation was 0 LOG (ivector-compute-dot-products[5.5.1697~1-22ac]:main():ivector-compute-dot-products.cc:111) Processed 80 trials 0 had errors.
@danpovey danpovey merged commit ee517cd into kaldi-asr:pybind11 Feb 4, 2020
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

5 participants