Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

The previous code is 30% slower than the baseline. #2

Draft
wants to merge 1 commit into
base: async-decoder
Choose a base branch
from
Draft

The previous code is 30% slower than the baseline. #2

wants to merge 1 commit into from

Commits on Feb 28, 2019

  1. The previous code is 30% slower than the baseline. 

    Improvements:
1. with the same config, the previous code will result in 10% more tokens in each frame compared with baseline. The reason is because GetCutoff function is over emitting tokens, while for baseline it is over emitting&non emiting tokens. (please check it use --verbose 6 and the number of all tokens is shown in line 895, but not in line 693, which only contains emitting) Hence we should use 10% less max_active config.
2. the baseline use kaldi version HashList, while the previous code use std::unordered_map. please still use HashList or its variants. The reason is because HashList is allocating memory in Pool/Block fashion, hence it is 10% faster than std::unordered_map. To show this problem, I temporally do a hack in line 247, plese check it

After these two improvements, the current code is with similar speed as the baseline. HOWEVER, please check the quality of 1-best & lattice, I have discuss about my worries here kaldi-asr#3061 (review)
    Zhehuai Chen committed Feb 28, 2019
    Configuration menu
    Copy the full SHA
    7e13314 View commit details
    Browse the repository at this point in the history