Closed yjcaimeow closed 1 year ago
Hi Yingjie, are you using this file to launch prior training?
The cfg.yml
for prior is different from the first-stage, but we will load the first-stage cfg.yml
through --config
(as in the train_prior.sh
)
Alternatively, I upload the prior config under the config
folder now. Note that the sde.vae_checkpoint
need to change to your stage-1 vae path.
Hi ZENGXH,
Thank you for your quick and patient reply.
Yes, I am training the second stage use "script/train_prior.sh". And I set the --config with the cfg.yml from first-stage vae folder. I understand you mean this experiment is okay, huh? If correct, I will keep training. By the way, is it convenient for you to share a loss curve of the second stage of training? The loss on my side is kept at around 1.9.
Thanks for the uploaded prior config and I will use it following experiments.
Best regards, Yingjie
Re 'experiment is okay': How do you fix the beta
shape error? Is it caused by wrong config?
Re loss curve: this is my curve for stage-2 training (on car): also the visualization of the reconstruction: visualization of samples at different steps (top is output points, bottom is latent poitns):
Hi ZENGXH,
Thank you for your quick and patient reply. The beta from the first-stage cfg is wrong. And I retrain the second stage with the cfg you uploaded recently. The curve and final results are right.
Thanks again.
Best, Yingjie
@ZENGXH I would like to second that the cfg generated by the train_vae.sh script does not work. I am getting the same beta issue as described at the top by @yjcaimeow.
Hi @ZENGXH ,
I successfully completed the first stage hvae training, but there is no cfg file when training the second diffusion prior. 1) Could you please provide the diffusion prior cfg file? 2) Is it okay to use the first stage vae cfg.yml for the second diffusion prior training process? I have tried, but the beta shape maybe wrong.
Best regards, Yingjie