emi-group / fapn Goto Github PK

Could you please explain what is the reason for multiplying feat_up by 2 here https://github.com/EMI-Group/FaPN/blob/main/detectron2/modeling/backbone/fan.py#L66 ?

Hi, I follow your instructions to arrange the project. However, this error still occurs. Could you help me with this issue?

KeyError: "No object named 'build_resnet_fan_backbone' found in 'BACKBONE' registry!"

Hello, I am very interested in your FaPN model. Seeing that the FPS of the COCO dataset is very high in semantic segmentation, can you give me this model? I want to reproduce it.

Hi, wonder:

1. Is that possible replace dcnv2 plugin with torchvision dcn API?
2. Can dcnv2 replaced with normal conv since dcn is not well supported in production. Also, will the performance drop a lot using normal conv?

Hi, thank you for your contribution. It is a good work.

I have a small question. Have you done any ablation studies about other FPN-like modules? i.e. FPN-NAP, PAN, BIFPN, etc. It will be much more convincable if you provide such result.

Thank you.

Hi, dear authors, can you release your codes for feature alignment visualization, i.e., codes for Figure 6 ?

没有找到fan.py中的 detectron2.layer文件

When putting the DCNv2 provided by FaPN on my own project, I met the problem of "RuntimeError: cublas runtime error : the GPU program failed to execute at /pytorch/aten/src/THC/THCBlas.cu:309". I used the cuda 11.0 and pytorch 1.7, and run on RTX4090, can you help me find the solutions?

Can this method be used in the unit price segment target detection model?

I want to use fapn structure in my code, but " File "/databank/home/DCNv2/dcn_v2.py", line 43, in forward
ctx.deformable_groups,
RuntimeError: expected scalar type Float but found Half" appears，How can I solve it? My version is cuda10.1, python3.7,pytoch 1.7.0

self.dcpack_L2 = dcn_v2(out_nc, out_nc, 3, stride=1, padding=1, dilation=1, deformable_groups=8, extra_offset_mask=True)
why out_nc is 256, not 216(3 x kernel_size x kernal_size x deformable_groups) ?

I was interested in your FaPN model, could you only provide FaPN structure?

Could you share the codes for feature visualization in your paper?

In module FeatureAlign_v2
offset = self.offset(torch.cat([feat_arm, feat_up * 2], dim=1)) # concat for offset by compute the dif
Why multiple feat_up by 2?

您好，看了您的论文，希望借助一下您论文中的FSM和FAM模块，直接插入到我的网络中，但我没有找到这两个模块的单独实现，请问这两个模块的单独实现在哪里呢，谢谢

已解决