How to build Caffe (Caffe 1.0)

Get source

• Get Caffe 1.0 source from: https://github.com/BVLC/caffe/archive/1.0.tar.gz

Edit Makefile.config

• Before

  # CPU_ONLY := 1
  
  # OPENCV_VERSION := 3
  
  # CUSTOM_CXX := g++
  
  BLAS := atlas
  
  INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include
  LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
  

• After

  CPU_ONLY := 1
  
  OPENCV_VERSION := 3
  
  CUSTOM_CXX := arm-linux-gnueabihf-g++
  
  BLAS := open
  
  INCLUDE_DIRS := <Path to zynq-library repo>/dl-framework/caffe-dependency/include <Path to zynq-library repo>/imaging/opencv/include
  LIBRARY_DIRS := <Path to zynq-library repo>/dl-framework/caffe-dependency/lib <Path to zynq-library repo>/imaging/opencv/lib
  

Build

• Be sure to source Petalinux's setting.sh

  $ PATH=<Path to protoc for host>/bin:${PATH} make -j`nproc`
  $ make distribute

---

Run example on Zynq

• Try MNIST training

  $ caffe.bin train --solver=examples/mnist/lenet_solver.prototxt
  
  I0720 06:00:32.205727  1070 caffe.cpp:211] Use CPU.
  I0720 06:00:32.207193  1070 solver.cpp:44] Initializing solver from parameters:
  test_iter: 100
  test_interval: 500
  base_lr: 0.01
  
  ...
  
  I0720 08:16:14.347036  1070 solver.cpp:310] Iteration 10000, loss = 0.00200753
  I0720 08:16:14.347198  1070 solver.cpp:330] Iteration 10000, Testing net (#0)
  I0720 08:16:55.999289  1073 data_layer.cpp:73] Restarting data prefetching from start.
  I0720 08:16:57.729313  1070 solver.cpp:397]     Test net output #0: accuracy = 0.9904
  I0720 08:16:57.729506  1070 solver.cpp:397]     Test net output #1: loss = 0.0299617 (* 1 = 0.0299617 loss)
  I0720 08:16:57.729576  1070 solver.cpp:315] Optimization Done.
  I0720 08:16:57.729624  1070 caffe.cpp:259] Optimization Done.