how much epoch I need to get the realistic results?
elsobhano opened this issue · comments
Hi everybody
I trained my network for 30 epochs. Is it normal that I don't have good results?
It could be the case. If possible, just try running for more epochs and compare results. Also, you can use checkpoints to verify that the model is still improving.
It depends on your data. Intuitively, however, >100 epoch is more likely to give promising result. I've tried pix2pix on my own dataset and it starts to generate good result after 100 epoch. Also you could try learning rate decay.
It depends on many factors: the number of images in your dataset, the complexity of your input/output domains, the task itself, the image resolution, etc. 30 epochs might not be enough unless you have lots of images in your dataset.
Hi all
my training take long time to get the result 1 hour per epoch, is it normal ?
note: the input image size 512x512
can i stop the training in any step i want??
@Sarmadfismael try reducing the image size (using crop or other options listed in the Training/test tips section). Also, it will depend on the memory available on your setup.
As far as I understood, there is an option to resume training if needed. So, yes you can stop and resume it. Search for the --continue_train flag in the Training/Test tips page.