How to train on own dataset?

[AnaZou]

Hi, i'm trying to train on my own dataset with different classes, when i revised the train.prototxt, i just revised three place[new conv layer:num_output; input_data:num_classes],is that right?

[YuwenXiong]

Hi @zxy1016 ,
Besides prepare your own data, you need to modify two layers' num_output: "rfcn_cls", "psroipooled_cls_rois", change the num_output of "rfcn_cls" to [your class_num] * 7 * 7, and the num_output of "psroipooled_cls_rois" to [your class_num]. And you needn't modify num_classes of "input-data" layer since it is useless when using RPN.

[qianyeqiang]

@orpine
@orpine
But why have i trained it using my own dataset without changing "rfcn_cls" and "psroipooled_cls_rois"?(my categories are 2)
Does it have impact on accuacy?

[Timonzimm]

Hi,
Could you please enlighten me on the part about preparing the data? I couldn't find any help reading the Readme