`sample_fast` images are incoherent with logged `no_pix` images during train.

[ysig]

Hi,

While training the Net2Net transformer model, after I have trained the VQ-GAN, I see interesting images being generated on the logs, with no_pix in their name, which probably means that they are sampled from scratch (without an image prior).

The problem is that when I then try to generate samples with sample_fast, the results are really, really bad, compared to the once in images/{train,val}/samples_nopix*.png of my model logs.

Why this incoherence? What I am doing wrong or understanding wrong? How can someone visualize the true progress, while training the model?

Thank you,

[hjq133]

I see when training, they use the function "sample", but in sample_fast.py, they use function "sample_with_past", this may cause the incoherence ? But I don't understand what sample_with_past is doing, is this a variant of normal transformer?

[hjq133]

btw, I still struggle with training the Net2Net transformer model, I can not get good sample images in my log. can you share the config files of yours? and how many images in your dataset, how many epochs it trains