search

PkuRainBow / HDC.caffe

Complete Code for "Hard-Aware-Deeply-Cascaded-Embedding"
GNU General Public License v3.0
97 stars 35 forks source link

About the train batch size #1

Open tendence opened 7 years ago

tendence commented 7 years ago

My Graphic card is GTX1080. So the batch size in my experiment cannot be set as large as 100. Will it decrease the accuracy?

PkuRainBow commented 7 years ago

Hi, I have not test other batch size~ I think you could test it and report the HDC method's result ~ I will be glad to see the influence of the batch. For example, if your batch size is 60, you could sample 10 classes with 6 images for each class. You should change the settings of the sample strategy.

zimenglan-sysu-512 commented 7 years ago

@PkuRainBow i found that changing the batchsize to 50 the loss won't decrease on CUB200_2011 dataset

PkuRainBow commented 7 years ago

@zimenglan-sysu-512 if you change the batch size. You are expected to change the sample strategies according to the batchsize.