search

zhanglonghao1992 / Deformable-ConvNets-v1-v2-caffe

Unofficial caffe implementation of Deformable ConvNet V1 and V2
31 stars 15 forks source link

something wrong in my code #2

Open lucinyaLi opened 5 years ago

lucinyaLi commented 5 years ago

hello, I have met some trouble in my code when I use the deformable convolution in SSD, the loss is always high and the mAP is very very low. I tried to change the deformable_group, but the solution has no change. So is there some trick in using the deformable convolution? And in the code, is the BN layer indispensable?

zhanglonghao1992 commented 5 years ago

I dont think I used any tricks training the deformable convnet. I suggest you print the “offsets” of every iter and see if they are reasonable.

lucinyaLi commented 5 years ago

I dont think I used any tricks training the deformable convnet. I suggest you print the “offsets” of every iter and see if they are reasonable.

Have you tried to train a net from random Initialization ？

zhanglonghao1992 commented 5 years ago

I dont think I used any tricks training the deformable convnet. I suggest you print the “offsets” of every iter and see if they are reasonable.

Have you tried to train a net from random Initialization ？

No