patrickltobing / cyclevae-vc-neuralvoco Goto Github PK

I am get strange CUDNN_STATUS_EXECUTION_FAILED error during VC model training.

OS: Docker container with Ubuntu 20.04
GPU: Two NVidia GeForce GTX Titan X
CUDA: 11.1
cuDNN installed, Nvidia tests passed OK.

No project installation (make) error.

Project: cycvae_mwdlp_vcc20

Data preparation passed OK.
Data preprocessing passed OK.

Set stage=4 (and comment others)
Set n_jobs=30 (and comment others)
Set GPU_device=0 (and comment others)
Set GPU_device_str=0 (and comment others)
Set n_gpus=1 (and comment others)

Run ./run.sh

After 1-2 minutes process aborted.

Log files attached.
cvnm_log.zip

Hello Patrick,

I noticed that when using the VCTK-Corpus the training fails and in the log file it shows:

  number of evaluation_data -- batch_size_eval = 0 -- 1

It seems that this occurs because in dataset.py on line 828 there is always a "p" in the file name and eval exist will remain false. What is the correct way to proceed?

Thanks,
Bob

Summary

Which version of implementation is equivalent to the paper model?

Main

Thanks for open-source great VC system.
I am trying to reproduce the result in CycVAE-MWDLP-VC paper.

Now there are v3.0 and v3.1 of implementation.
Which is the exact implementation of the paper?
v3.0 is "pre-publish" & v3.1 is "the paper" model?
or, v3.0 is "the paper" & v3.1 is "improved" model?

Again, thanks for great open-source VC system.

Hi, thanks for your great work!

When i run the run.sh with stage 4 in egs/cycvae_mwdlp_vcc20, core dumped occurred when training the voice conversion model. It seems like some generated hdf5 items caused the problem, so far the items I found are hdf5/tr_vcc20_10ms_24kHz/TEM2/E20019.h5 and hdf5/tr_vcc20_10ms_24kHz/TMM1/M10029.h5.

After replacing the hdf5/tr_vcc20_10ms_24kHz/TEM2/E20019.h5 with another normal one, the training can go on. But core dumped still occurs when I replace hdf5/tr_vcc20_10ms_24kHz/TMM1/M10029.h5 with another hdf5 file.

Can you help me to check where is wrong? Thanks for your time!!

Hi

I'm trying to fine tune the cyclevae model with my own data and I'm using the pretrained model of the compatible demo. I've done all the stages of preprocessing (0init123) and I don't need to do the stages 4 and 5. In stage 6 when I pass the pkl file to the script it requires another checkpoint file (checkpoint 53). I renamed the checkpoint file to checkpoint 53 and I got torch size mismatch error.

Can you tell me how to fine tune the compatible demo with my own data? Thanks for your reply.