Quesation finetune a custom dataset with only 33 class

[jerryho-quanta]

Hi sfzhang15,

I am interested in fine-tune part in your git. I found you fine-tune form COCO to VOC, and you don't freeze the VGG part. You use coco pre-trained caffe to re-train all mode (VGG+Refinedet)? If I want to use coco/voc weight to fine-tune my only 33 classed with about 1000 images, should I freeze the backbone part(VGG16)?

Thanks,

Jerry Ho