search

MiukkaZh / MGT

Learning Domain-Invariant Transformation for Speaker Verification.
9 stars 0 forks source link

trials #1

Open wangsheng3 opened 1 year ago

wangsheng3 commented 1 year ago

During training, it was found that some of the pkl files listed in the trials file you provided were not present. For example, in one of your trials files, you need to use test/id00870-rama-02-002.pkl, but the file extracted from the dataset does not contain this file. What is the problem?

MiukkaZh commented 1 year ago

The CN-Celeb dataset has been updated in the past. You can find the previous versions of the CN-Celeb dataset, or generate trials for the current version of CN-Celeb. The code cnceleb_trails.py for generating files and trials has been uploaded to GitHub. You can refer to this.

wangsheng3 commented 1 year ago

Hello, I have studied your newly uploaded python file, but there is one thing I don't quite understand: lines 7 and 21 in the train_task function, and lines 101 and 116 in the fix_trial function. Aren't these two parts inconsistent?

MiukkaZh commented 1 year ago

Due to the previous processing of trials, we have obtained large-scale trials covering different genres, namely fix_cnceleb_trials, based on trials.lst. Therefore, in the subsequent train_task, it can be directly processed using it as a foundation. You can directly run it on trials.lst, as shown in Line 8. The relevant sections have been updated on GitHub.