NonConverge

[tyut11103]

Hello Mr maximus.mutschler:
I have read your paper carefully and very interest in your work. The PAL agorithm you provided can automatically find a good learning rate without the need of hand-designed step size schedules.
I want reproduce the experiment of Resnet50 on ImageNet , BUT the training loss did not converge. The optimizer used the PalOptimizer come from your pytorch code ,and tried the recommended hyperparameters. Can I get your hyperparameter configuration?
Waiting for your reply. Best Regards