Learning Rate Decay Method

[kuan-wang]

Hello, That's really a good result :) Can you share your training method to train mobilenetv2 that high?( such as default learning rate and weight-decay method) . Thanks~
I am currently training the model with SGD and keeping other hyper-parameters the same (except that I use batch size 256), but only get Top1: 71.17

[ericsun99]

I have training with SGD, and have using some data augment method.

[SophieZhou]

@ericsun99 which data augmentation did you use during your training? Using crop, flip, resize, I only got about top 1 accuracy 68 % and top5 about 88%.
And I evaluate the model by center crop.

[SophieZhou]

@ericsun99 @THUKey In your test, do you test the model with tencrops?

[Coderx7]

@THUKey ,@ericsun99 could you please share your hyperparameters?

[CF2220160244]

Hello @THUKey . I use SGD train with a 1080Ti, batchsize=96, lr=0.045, weight decay=0.00004, and decrease the lr 0.98 for each epoch, after 2 days only get 67%, can you tell me your GPU, batchsize, and the initial lr to help me.
I am a student in beijing institute of technology. Thank you very much!