yeephycho/nasnet-tensorflow

How many epochs during finetuning ?

phongnhhn92 opened this issue · 4 comments

phongnhhn92 commented

Hello sirs, I have one question. I can not find the information related to the number of epochs you have used to finetuning flowers dataset with NASnet. I tried to read the code but there are no information either. Can you clarify it ?

yeephycho commented

Hi @phongnhhn92 ,

There's no need to fix the epoch, you can stop the training when the the validation accuracy stops improve. Usually 30 epoch to 90 epoch will be fine depends on your learning rate and decay policy.

With regards!
Yeephycho

phongnhhn92 commented

Thanks for your answer ! But there is one thing I dont understand. Consuming that i am using the original config for finetuning (batch size 32, optimizer rmsprop) and I stop my training at step 50k, then how to know how many epochs did I perform finetuning ? Sorry for my silly question if it is.

yeephycho commented

Training_step * batch_size / total number of training image = epoch

phongnhhn92 commented

Thanks for your help !