py-faster-rcnn on Ubuntu 17.10 and Cuda 9.0

faster-rcnn-caffe-master.md

Instructions

• install cuda-9.0 (https://developer.nvidia.com/cuda-toolkit-archive)
• install cudnn-7 (https://developer.nvidia.com/cudnn)
• install dependencies for caffe
  • see the caffe installation guide [1] and wiki page [2]
• install boost-1.65 or higher (http://www.boost.org/users/history/version_1_65_1.html) [3]
• update caffe-fast-rcnn
  • Easiest way is to use a rebased branch on BVLC/master. See the notes below if you would rather merge the changes yourself.

  $ cd $FRCN_ROOT/caffe-fast-rcnn
  $ git remote add acmiyaguchi https://github.com/acmiyaguchi/caffe-fast-rcnn.git
  $ git checkout acmiyaguchi/faster-rcnn-rebased
  

  • optional: rebase against upstream master

  $ git remote add caffe https://github.com/BVLC/caffe.git
  $ git rebase caffe/master
  

• set variables in Makefile.config
  • USE_CUDNN := 1
  • WITH_PYTHON_LAYER := 1
  • OPENCV_VERSION := 3
  • CUSTOM_CXX := g++-6
    • I symlink gcc-6 and g++-6 to gcc and g++ in /usr/local/cuda-9-0/bin/ instead of this setting, but either should work
  • remove compute_20 from CUDA_ARCH since it's been depreciated in cuda-9-0
  • add hdf5 for linking since it's been renamed in 17.10
    • INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/include/hdf5/serial/
    • LIBRARY_DIRS := $(PYTHON_LIB) /usr/lib/x86_64-linux-gnu/hdf5/serial/
• build caffe
  • make -j${N_THREADS} all && make pycaffe
• change self.param_str_ to self.param_str in $FRCN_ROOT/lib/rpn/proposal_layer.py.
  • see rebased-caffe.patch below for the full diff
• validate that ./tools/demo.py works

Notes on Rebasing

BVLC/caffe and rbgirshick/caffe have diverged since this workaround [4] in March 2017. I've rebased the faster-rcnn patchset against the master caffe branch for convenience. [5]

This branch should merge cleanly into BVLC/caffe with the exception of the license.

If you rather merge the upstream changes yourself, the patch in [6] should outline the changes you need afterwards. -X theirs may not be the best merge policy. In particular, changes to the src/caffe/caffe.proto schema will overwrite the ROIPoolingParameter.

rebased-caffe.patch

diff --git a/lib/rpn/proposal_layer.py b/lib/rpn/proposal_layer.py
index b157160..e6d28cc 100644
--- a/lib/rpn/proposal_layer.py
+++ b/lib/rpn/proposal_layer.py
@@ -23,7 +23,7 @@ class ProposalLayer(caffe.Layer):
 
     def setup(self, bottom, top):
         # parse the layer parameter string, which must be valid YAML
-        layer_params = yaml.load(self.param_str_)
+        layer_params = yaml.load(self.param_str)
 
         self._feat_stride = layer_params['feat_stride']
         anchor_scales = layer_params.get('scales', (8, 16, 32))

• [1] http://caffe.berkeleyvision.org/installation.html
• [2] https://github.com/BVLC/caffe/wiki/Ubuntu-16.04-or-15.10-Installation-Guide
• [3] ubuntu 17.10 currently ships with boost 1.62. This uses the macro __CUDACC_VER__ that was depreciated in cuda-9.0. The current distribution on the boost site fixes this. See the bug tracker for more details. https://svn.boost.org/trac10/ticket/13152
• [4] this currently requires quite a bit of manual conflict resolution https://github.com/rbgirshick/py-faster-rcnn/issues/509?_pjax=%23js-repo-pjax-container#issuecomment-284133868]
• [5] https://github.com/acmiyaguchi/caffe-fast-rcnn/commits/faster-rcnn-rebased
• [6] https://github.com/acmiyaguchi/caffe-fast-rcnn/commit/d58e95906f1f1ba3557e3f46c57fa6a4a1361be2

@acmiyaguchi Thank you very much. This report helps me to solve my problems.