gomxnet

Amalgamation and go binding for mxnet https://github.com/dmlc/mxnet

Go binding for predictor

• Use go get github.com/jdeng/gomxnet to install. If your openblas version happens to be 0.2.14_1 you are all set.
• Update predict.go: update openblas library path accordingly (the first two lines).

...
//#cgo CXXFLAGS: -std=c++11 -I/usr/local/Cellar/openblas/0.2.14_1/include
//#cgo LDFLAGS: -L /usr/local/Cellar/openblas/0.2.14_1/lib/ -lopenblas
//#include <stdlib.h>
//#include "mxnet.h"
import "C"
import "unsafe"
...

• go build could be slow when importing this library due to recompiling mxnet.cc every time. One workaround is manually building and linking mxnet.a. You will need to remove mxnet.cc to prevent go build compiling it. See below

cd $GOPATH/src/github.com/jdeng/gomxnet/amalgamation
sh build.sh
ls mxnet.a 
rm ../mxnet.cc

And add one line to predict.go. It needs to be an absolute path.

//#cgo LDFLAGS: /<path-to-gopath>/src/github.com/jdeng/gomxnet/amalgamation/mxnet.a -lstdc++

• Build the sample example/main.go with go build in the example directory. You need to install a dependent library with go get github.com/disintegration/imaging. ** Download the model file package from [https://github.com/dmlc/mxnet-model-gallery] and update the path in main.go. Build with go build. Try with ./example cat15.jpg, the program should be able to recognize the cat.
• Tested with golang 1.5.1 on Mac OS X Yosemite.
• Sample usage (from src/main.go)

  // read model files into memory
  symbol, _ := ioutil.ReadFile("./Inception-symbol.json")
  params, _ := ioutil.ReadFile("./Inception-0009.params")
  
  // create predictor with model, device and input node config
  batch := 1
  pred, _ := gomxnet.NewPredictor(gomxnet.Model{symbol, params}, gomxnet.Device{gomxnet.CPU, 0}, []gomxnet.InputNode{{"data", []uint32{batch, 3, 224, 224}}})

  // get input vector from 224 * 224 image(s)
  input, _ := gomxnet.InputFrom([]image.Image{img}, gomxnet.ImageMean{117.0, 117.0, 117.0})
  
  // feed forward
  pred.Forward("data", input)
  
  // get the first output node. length for each image is len(output) / batch
  output, _ := pred.GetOutput(0)
  
  // free the predictor
  pred.Free()

mxnet amalgamation (this is optional. A pre-generated mxnet.cc is already in the directory.)

• Check out mxnet, e.g., in ~/Sources/, update submodules and build
• Generate mxnet.cc in amalgamation directory using amalgamation/gen.sh (content shown below). You may need to update the first two lines to point to your mxnet and openblas locations.

export MXNET_ROOT=~/Source/mxnet
export OPENBLAS_ROOT=/usr/local/Cellar/openblas/0.2.14_1
rm -f ./mxnet
echo "Linking $MXNET_ROOT to ./mxnet"
ln -s $MXNET_ROOT ./mxnet
echo "Generating deps from $MXNET_ROOT to mxnet0.d with mxnet0.cc"
g++ -MD -MF mxnet0.d -std=c++11 -Wall -I ./mxnet/ -I ./mxnet/mshadow/ -I ./mxnet/dmlc-core/include -I ./mxnet/include -I$OPENBLAS_ROOT/include -c  mxnet0.cc

echo "Generating amalgamation to mxnet.cc"
python ./expand.py

cp mxnet.cc ../
echo "Done"

TODO

• Merge openblas into the amalgamation?
• Add Train API?