Retraining NCSN++ using PDS

[Aravinda27]

Hi Hengyuan

Let me first congratulate for the nice work. Can we use the Preconditioned diffusion sampling (PDS) for training NCSN++ from scratch. The reason is because I modified the time step embeddings of NCSN++ to Resnet based embeddings. So I need to retrain the NCSN++ from the scratch. What do you suggest??

[AwakerMhy]

Thank you for your attention to our work. PDS is applied during the sampling phase of NCSN++. If you want to train the NCSN++ using PDS, our suggestion is that you may need to reconsider the filters, since our PDS will change the objective function, leading to unexpected results, such as instability or harder convergence of training.

[Aravinda27]

Thank you for your reply.

[Aravinda27]

Hi Hengyuan,

Sorry for troubling again where is this time_cond parameter is calculated in ncsnpp.py file. Since it is used in the forward function of ncsnpp.py file.

[AwakerMhy]

This parameter is set by the step schedule which is passed to the update_fn of a corrector or predictor iteration function in sampling.py, please take a look.