mingjiechen / lowresourcevc Goto Github PK

Firstly , I found that you have been implemented the code of style encoder contrastive loss , why do not you use it ?
Secondly , I found that the gradient penalty can make convergence faster , why do not you use it ?
I wish you can answer me . Thank you very much.

Nice job, @MingjieChen ,

and i'm so interested in ur arxiv paper , i will read it later.

1. I wanna figure out how u train the speaker encoder in the process, did it pretrained before we train the VC model? Or u choose to train together at the same time?

2. I noticed that u described in ur paper that the embedding is 126 dims, and u still wanna use the CIN model? in the original CIN paper in image field, it pointed out that the related embedding use [1, feature_classes]x[feature_classes, feature_dim], and in that way we calculate the \gamma & \beta, so as to make current embedding have more relationship between all speakers. See more from the relates issues about CIN in image_field
   Maybe my thought is wrong, i wanna know how the embedding that not in the form of one-hot that can use CIN patern, and use \gamma & \beta to calculate the related embedding?

3. And would u please write the preparation processes of data preparing? I wanna have it a try~~~

All the best,
Luke Huang