SkillAgentSearch skills...

Mxnet2caffe

No description available

Install / Use

/learn @DG-Apollo/Mxnet2caffe
About this skill

Quality Score

0/100

Supported Platforms

Universal

README

Mxnet2Caffe: A tool to convert model from MXNet to Caffe

  • Modern C++ Practice
  • High performance
  • No dependency execpt for Caffe

Required Packages

BVLC/Caffe

You must clearly know the absolute path where your caffe installed. If your caffe is believed to have installed in YOUR_CAFFE_HOME, please try to verify those caffe files do really exsist:

ls YOUR_CAFFE_HOME/distribute/include/caffe/caffe.hpp
ls YOUR_CAFFE_HOME/distribute/lib/libcaffe.so

or

ls YOUR_CAFFE_HOME/build/install/include/caffe/caffe.hpp
ls YOUR_CAFFE_HOME/build/install/lib/libcaffe.so

Please DO NOT complains any building failure before you really did all above confirmations.

Build

mkdir build && cd build
cmake -DCAFFE_HOME=<YOUR_CAFFE_HOME> ..
make # or make -j8

To Prepare a MxNet Model

A MxNet model consist of a symbol file (*.json) and a parameters file (*.params). In general settings, the two files should have name like:

  • <model_name>-symbol.json and
  • <model_name>-0000.params.

Mxnet2Caffe use an independent config json file to describe your model and to configure conversion settings. Following is an example of the config json file:

{
	"mxnet_json" : "squeezenet_v1.1-symbol.json",
	"mxnet_params" : "squeezenet_v1.1-0000.params",
	"caffe_prototxt" : "squeezenet_v1.1.prototxt",
	"caffe_caffemodel" : "squeezenet_v1.1.caffemodel",
	"inputs" : [
			{"name" : "data", "shape" : [1, 3, 224, 224]},
			{"name" : "prob_label", "shape" : [1]}
		]
}

Find more examples in sub-path ./examples.

Properties used by the config json:

It should be very clear in the above example.

Running the conversion:

Simply run command ./mxnet2caffe config.json and a Caffe model will be presented after conversion by your configurations.

View on GitHub
GitHub Stars6
CategoryDevelopment
Updated6y ago
Forks3

Languages

C++

Security Score

50/100

Audited on Nov 4, 2019

No findings