Closed zh794390558 closed 2 years ago
Sorry, I am still busy now for this weekend. Sugget you to see subtools/recipe/voxceleb/gather_results_from_epochs.sh to know the example of score normalization. For score fusion, see subtools/weightScore.sh, subtools/fusionByLda.sh or subtools/fusitonBySvm.py to have a try. Note that, the format of trials is [target-id utt-id target/nontarget] and score is [target-id utt-id score].
I will try and anytime is ok.
using svm fusion, the result is minDCF is 1.0000 at threshold 1.0000 (p-target=0.05, c-miss=1,c-fa=1)
how to debug this?
Hi, could you try the fusion with equal weights? Make sure it is no problem and then use LDA fusion to have a try. For SVM fusion, you should save the weights firstly and then apply these to your scores. By the way, you could also check the performance with EER metric.
On Aug 12, 2020, at 9:31 PM, Hui Zhang notifications@github.com wrote:
using svm fusion, the result is minDCF is 1.0000 at threshold 1.0000 (p-target=0.05, c-miss=1,c-fa=1) how to debug this?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
fusion with equal weights is wok, and I have not try LDA now, SVM fusion is not work for me.
Hi, are there negative weights? Try to plot a DET curve to observe the whole status.
Sent from my iPhone
On Aug 17, 2020, at 11:30 AM, Hui Zhang notifications@github.com wrote:
fusion with equal weights is wok, and I have not try LDA now, SVM fusion is not work for me.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
why https://github.com/Snowdar/asv-subtools/blob/master/fusionBySvm.py#L142 txt_b
is zero?
does score norm always improve EER and minDCF?
Hi, it depends on the cohort set and you could not always select a good cohort.
On Sep 8, 2020, at 4:23 PM, Hui Zhang notifications@github.com wrote:
does score norm always improve EER and minDCF?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
Yeah, thanks to your suggestion and I will write this in my free time (about this weekend).