Have you tried to train from scratch and use large learning rate?

[ssssholmes]

The authors mentioned they use adamw with 0.003 learning rate without imagenet pretrain. I tried it and find it is really hard to train with such config.

[gaopengcuhk]

The largest learning rate for stable training in this codebase is 0.001.

[Sharpiless]

Thanks for your great work. But the paper says:"We use the stochastic depth (Huang et al., 2016) with a rate of 10% to reduce overfitting." Have you tried to use stochastic depth method in your model?

[gaopengcuhk]

I will try stochastic depth in the future.