Training of the factorized model

[ildmxGL]

Hi, @fracape

I trained factorized model for 100 epochs, but inference results for kodak dataset were not similar.

"kodak-{number}" means just a model number. The former is the 3.0M model, and the latter is the 7.0M model. The lambda is 0.01; aims quality 3.

Are these vary results usual? Do I have to train several models for best results as the pre-trained?
Or did the models just not converge?

Thanks for your great works.

[fracape]

Hi, sorry for the late response, I missed this. If still relevant, which training set did you use? Did you monitor your losses? Your bitrates are very weird.