py faster rcnn + cudNN V2

[kejriwalnishant1990]

Are you able to run py faster rcnn with cudNN v2 (which is officially supported on TK1)?

[joeking11829]

@kejriwalnishant1990

Based on my memory. I use it with cuDNN v2 on TK1.

[kejriwalnishant1990]

@joeking11829 Well, I am able to run it without cudNN. Can you please tell me caffe version and other things which you have changed to run it with cudNN v2? Thanks for the reply.

[joeking11829]

@kejriwalnishant1990

What's the Error Message when you run it with cuDNN v2.

[kejriwalnishant1990]

@joeking11829 The recent caffe does not support cudNN v2 and it throws fatal error that it needs cudNN v3 or more. I commented that section of finding cudNN v3, and but then it throws error in cudNN functions definition because there are lot of functionalities modification from v2 to v3. So I want to know whether did u make any changes in caffe? Which version of caffe you have used for Faster RCNN?

[kejriwalnishant1990]

The error message after commenting cudNN v3 fatal error message:

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(14): error: identifier "normdesc" is undefined

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(14): error: identifier "CUDNN_DIVNORM_PRECOMPUTED_MEANS" is undefined

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(14): error: identifier "cudnnDivisiveNormalizationForward" is undefined detected during instantiation of "void caffe::CuDNNLCNLayer::Forward_gpu(const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &, const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &) [with Dtype=float]" (43): here

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(32): error: identifier "normdesc" is undefined

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(32): error: identifier "CUDNN_DIVNORM_PRECOMPUTED_MEANS" is undefined

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(32): error: identifier "cudnnDivisiveNormalizationBackward" is undefined detected during instantiation of "void caffe::CuDNNLCNLayer::Backward_gpu(const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &, const std::vector<__nv_bool, std::allocator<__nv_bool>> &, const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &) [with Dtype=float]" (43): here

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(14): error: identifier "cudnnDivisiveNormalizationForward" is undefined detected during instantiation of "void caffe::CuDNNLCNLayer::Forward_gpu(const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &, const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &) [with Dtype=double]" (43): here

/home/ubuntu/caffe-tk1-new/src/caffe/layers/cudnn_lcn_layer.cu(32): error: identifier "cudnnDivisiveNormalizationBackward" is undefined detected during instantiation of "void caffe::CuDNNLCNLayer::Backward_gpu(const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &, const std::vector<__nv_bool, std::allocator<__nv_bool>> &, const std::vectorcaffe::Blob<Dtype , std::allocatorcaffe::Blob<Dtype >> &) [with Dtype=double]" (43): here

8 errors detected in the compilation of "/tmp/tmpxft_00001555_00000000-6_cudnn_lcn_layer.cpp1.ii". CMake Error at cuda_compile_generated_cudnn_lcn_layer.cu.o.cmake:264 (message): Error generating file /home/ubuntu/caffe-tk1-new/build/src/caffe/CMakeFiles/cuda_compile.dir/layers/./cuda_compile_generated_cudnn_lcn_layer.cu.o

make[2]: * [src/caffe/CMakeFiles/cuda_compile.dir/layers/./cuda_compile_generated_cudnn_lcn_layer.cu.o] Error 1 make[1]: * [src/caffe/CMakeFiles/caffe.dir/all] Error 2

[joeking11829]

I use the specific commit of "caffe-fast-rcnn @ 4115385" with CuDNNv2 only.