search

YuwenXiong / py-R-FCN

R-FCN with joint training and python support
MIT License
1.05k stars 472 forks source link

How to implement R-FCN using ZF or VGG base model #100

Open liuliu66 opened 6 years ago

liuliu66 commented 6 years ago

I tried to implement R-FCN for my own dataset. But my dataset contains lots of tiny objects and I think a relative shallow base network could be better for tiny objects. So I want to train R-FCN using ZF or VGG as base model but I could not find prototxt file for ZF in R-FCN and only faster rcnn. Could anyone help me to implement it?

RebornL commented 6 years ago

@liuliu66 Have you finish it ? I use VGG_M model to train, but I find that it can't save model. Maybe the prototxt from faster-rcnn is not correct. The error message show below: File "/home/yan/disk1/phone_rfcn/py-R-FCN-phone_v1.2/tools/train_net.py", line 119, in <module> max_iters=args.max_iters) File "/home/yan/disk1/phone_rfcn/py-R-FCN-phone_v1.2/tools/../lib/fast_rcnn/train.py", line 184, in train_net model_paths = sw.train_model(max_iters) File "/home/yan/disk1/phone_rfcn/py-R-FCN-phone_v1.2/tools/../lib/fast_rcnn/train.py", line 132, in train_model model_paths.append(self.snapshot()) File "/home/yan/disk1/phone_rfcn/py-R-FCN-phone_v1.2/tools/../lib/fast_rcnn/train.py", line 77, in snapshot self.bbox_stds[:, np.newaxis]) ValueError: operands could not be broadcast together with shapes (20,1024) (8,1) Do you have any idea? Thanks!