-
Notifications
You must be signed in to change notification settings - Fork 55
Open
Description
Hi @camel007 ,
ShuffleNet is a great idea and the model size is apparently reduced.
Have you ever tried to train it on imagenet?
I only got 64% as top1 acc (1x, group = 3). My training schedule is 100 epochs, with batchsize 256, lr rate starts from 0.1 and decay by 0.1 every 30 epochs.
I'd be appreciated if there's any tricks in training to reach the paper acc.
thx!
Metadata
Metadata
Assignees
Labels
No labels