Training log

[XiSHEN0220]

Dear authors,

Thanks a lot for sharing the code, which is quite clean and easy to read !!

I plan to train the model from scratch.
As the training takes ~10 days (c.f. issue 11), I am wondering whether it is possible to share your training log corresponding to the experiments of Lsun churches (VQGAN and Absorbing Diffusion sampler).

Finally, in the paper Sec.4.4, you mentioned that leveraging Diff Aug + $\lambda_{max} = 1$ leads to better performance. However, in the provided config, if I understand correctly, the diff aug is set to False.

[XiSHEN0220]

Dear, authors,

I have another question about the FID evaluation.

If I understand correctly, the FID provided in the code is computed on the training set? (i.e. real and reconstructed images are both for the training set)

The details can be found in the following scripts:

• calc_VQGAN_FID line 18
• vqgan_utils line 78
• vqgan_utils line 93-111
• data_utils line 152-193

If this is true, does it make sense to report FID on the training set?

Thanks for your time !!
Best,

[peterhessey]

Hi, thanks for your interest in our paper and code! I shall answer your questions in order:

I am wondering whether it is possible to share your training log corresponding to the experiments of Lsun churches

• Unfortunately we can't share training logs currently as they contain references to sensitive university info that would need to be cleaned out first. If this would be of significant use to all users though we may be able to facilitate this in the future.

However, in the provided config, if I understand correctly, the diff aug is set to False.

• Good spot! That should not be set to False by default, please change it if you're trying to repeat experiments :)

If this is true, does it make sense to report FID on the training set?

It makes sense in this setting because in our experiments we cared about evaluating raw reconstruction quality, which is impacted by training stability and how many codes are used, among other issues - what we were trying to address with our VQGAN changes, not generalisation. As a side note, in later experiments we found that DiffAug helps with generalisation quite a bit.

[XiSHEN0220]

Thanks for your reply !!!
It really helps me a lot~~
I close the issue.