Hi,
Are you realized what was the problem ? I am trying to fine-tune this model on a dataset of 300k images. When I use the whole dataset, the model stops improving, and the loss only oscillates between 0.10 and 0.12. I’ve trained the model for about 20k steps with a batch size of 32, but I can’t see any improvement after the first 1k steps. I test different learning rate from 3e-4 ,1e-5,5e-5,1e-5,1e-6. nothing happened.
here are some of my result:
A good result is middle. the side are not suitable and full of random particles. more than 80% of samples are not good like these two.