Some observations after 290k steps about freevc HOT 6 CLOSED

Btw, eval section of the tensorboard shows nothing. Is this expected?

from freevc.

Tested on my 290k checkpoint and it can convert properly. I guess your problem might be caused by dataset imbalance (i.e. data of some tgt speakers are too little compared to other speakers).
Eval section of the tensorboard is not empty. Maybe you did not tick on '<log_dir>/eval' on the bottom left corner of tensorboard.

from freevc.

Indeed my custom speakers have fewer total time per speaker (about 600-700 seconds) than VCTK speakers. But why male to male conversion works OK for the custom speakers.

I guess I have to level out the data for all speakers.

from freevc.

No idea. (＠_＠;)

from freevc.

In the original VCTK there are speakers ranging from 600 seconds to 1000 seconds. I'll try with the vocoder trained on the joint dataset and see whether this is the actual reason why conversion isn't consistent for additional speakers.

from freevc.

Btw, eval section of the tensorboard shows nothing. Is this expected?

@skol101 , You did haven't computed and wrote eval loss to tensorboard in train.py @OlaWod

from freevc.