search

miao0913 / SaCNN-CrowdCounting-Tencent_Youtu

Crowd Counting Via Scale-adaptive Convolutional Neural Network
107 stars 41 forks source link

batch_size of train and test #15

Closed zllljf closed 6 years ago

zllljf commented 6 years ago

in your paper, you said you set the batch size as 1 but if i want to change the batch_size to 4, i find i change the parameter of batch_size in train.prototxt is not work, if i should change other parameters in other script? thanks for your reply

miao0913 commented 6 years ago

Hi, I did remember using batch size more than 1 needs some specific setting (maybe due to memory or image size ? ). I did not see you error report, but I can give you two alternative options:

  1. make sure the images in the same batch are of the same size, which can be divided by 16 either W/H
  2. use two GPU to simulate batch size =2 I do not remember increasing batch size (or GPU number) gives out a big boost.
zllljf commented 6 years ago

thanks for reply maybe i wasn't clear enough. what i want to say is: i have croped the image into a size can dicided by 16, if train with 1 batch_size, the memory is used 1/4 , so i want to increase the batch_size of train. where i can find batch_size parameter is in train.prototxt, but i chang this parameter 1 to 4, the memory is used 1/4 too. so i think just change this parameter is not work, would like to tell me, where else do i need to change?

miao0913 commented 6 years ago

Can I see the error report ?