Open mawangyi opened 6 years ago
Current Deformable Convolution doesn't contain a CPU implementation.
@YuwenXiong thank you! if i want to implementation it use cpu,what i should to do?
@mawangyi I implemented the CPU version and have it running well.
For CPU based inference, res101 network takes 2-3 secs for a 800x450 resolution image. I just tested res50 that takes just 1.5 secs for the same 800x450 image. GPU-mode is typically 10 times faster than CPU.
Email me at arun.buduri@gmail.com and I can give you the CPU implementation. I fully don't know the process of submitting the code into mxnet repo (yet) but once I figure that out, I'd love to get that into everyone's hands.
@arunbuduri ,Thanks for the kindly reply! Could you please give me the CPU implementation code ? my email is mawangyi@163.com, Thanks&Best Regards!
Really sorry for the long delay in getting this information to you all.
I have captured the steps (along with the necessary source code) here - https://drive.google.com/open?id=1P8vYr9ekWp3NK8_N2bs8sl7jRT39m25I
I have tested this on Windows, Mac, Linux, and Ubuntu.
Please let me know if you have any questions.
Deformable-ConvNets is Cool, i need implementation it in cpu-mode。 so, i change the code: Replace GPU with CPU nms = cpu_nms_wrapper(config.TEST.NMS) context=[mx.cpu()]
now, i got an error: <NDArray 300x5 @cpu(0)> -------------DataParallelExecutorGroup--outputs---380---------------
the problem at “DataParallelExecutorGroup.py(387)get_outputs()” cpu can get “self.execs[0].outputs[1]”
anyone know how to Solve it???
Thanks a lot!