training epochs

[ZTYcj]

Thanks for your great work!
I have a question on the value of training epochs. I trained the "generate_flows_v2" with 2000 figures for 800 epochs each, and trained the "train_encoder_py" for 20000 epochs. The loss seems to have converged, but I can't get the good results like your pretrianed model. So how many epochs at least should I train according to your experience?
Thanks a lot!

[maria-larchenko]

Hi! We actually observed that for training dataset of flows we need around 3000 of images. Also I would recommend first to increase a number of training iteration for every flow to at least 20 000 (10k for the initial step and 10k for the rectified step). Do you work with B6 or B0 encoder? With B0 it should be possible to have a decent result with, I think, at least 50 000 epochs.