Closed ipod825 closed 4 years ago
@ipod825 I found the best model in the repo has val acc = 97. So I guess it is normal.
I met the same problem. As I run python main.py --is_train 1, the train acc converged to 91% and the val acc converged to 97% about 30 epochs later. But after 30 epoched, the train acc began to decrease and the loss began to increase. Could you tell me what is the problem, please? @ipod825@ifgovh
So, I can't understand:1)why the val acc is always higher than the train acc;2)why the train acc began to decrease and the loss began to increase after 30 epoches
I guess it is due to the stochasticity caused by reinforcement learning. Maybe the hybrid loss function does not converge well.
I have run python main.py --is_train 1 for several times,and the results are the same. What's more,he train acc converged to 91% and I haven't catch higher train acc.
I checkout the recent commit that changed the optimizer. Though you claim that "With the Adam optimizer, paper accuracy can be reached in 30 epochs." But as I run
python main.py --is_train 1
, the performance isn't so desirable as claimed. Here's the log of my running result. Can you confirm on this? (run based on commit 99c4cbe)