search

msracver / Deformable-ConvNets

Deformable Convolutional Networks
MIT License
4.04k stars 959 forks source link

Question DCNv2 operators #235

Closed Google1234 closed 5 years ago

Google1234 commented 5 years ago

Thanks for your great work! " The efficiency of processing multiple images in a mini-batch is considerably improved. In the previous operator, S is fixed as 1. In the updated operator, S can be set by the im2col_step parameter, whose default value is min(N, 64). The updated operator is significantly faster than the existing one when the image batch size is large."

Why not fix parameters S=1 ,to Process all inputs simultaneously ?Is it limited to GPU memory?

ancientmooner commented 5 years ago

The increased memory by larger S is minor. So it is okay to have larger value of S.

nian-liu commented 5 years ago

Hi @Google1234 , I am a newbie to Mxnet, could you tell me how to use the provided DCNv2 operators? I guess I need to copy the code in the DCNv2_op folder into the Mxnet source codes and then recompile the Mxnet? Could you give me some detailed hints? Thank you.