网络/光纤>> node-menoh>> 返回
项目作者： pfnet-research

项目描述 ：
NodeJS binding for Menoh DNN inference library
高级语言： JavaScript
项目地址: git://github.com/pfnet-research/node-menoh.git
创建时间： 2018-07-10T04:37:11Z
项目社区：https://github.com/pfnet-research/node-menoh
开源协议：MIT License
下载


menoh

NodeJS binding for Menoh DNN inference library.

Features

  • Fast DNN inference on Intel CPU.
  • Support standard ONNX format.
  • Easy to use.

Requirements

  • MKL-DNN library v0.14 or later.
  • ProtocolBuffers (Tested with v3.5.1)
  • Menoh(C/C++) library v1.x (Tested with v1.1.1)
  • NodeJS v6 or greater

Supported OS

  • Mac
  • Linux
  • Windows

Installation

Add menoh npm module to your project dependencies (package.json):

  1. npm install menoh -S

Installing dependencies

Mac & Linux

Simply follow the instruction described here.

For linux, you may need to add /usr/local/lib to LD_LIBRARY_PATH depending on your linux distrubtion.

  1. export LD_LIBRARY_PATH=/usr/local/lib

Or, you could add /usr/local/lib to system library path.

Windows

menoh import library and dll

You can download pre-build DLLs from here.
The import library (menoh.lib) and its header files are bundled in this module and built during
its installation.

Current version uses the import library built with native menoh v1.1.1.

Copy menoh.dll found in the pre-build package to C:Windows\System32\.

MKL-DNN

Follow this instruction to install MKL-DNN
lbrary and its dependencies. You may optionally download prebuild pacakge from here.

The mklml.dll (included in the pre-built package for the native menoh v1.1.1) depends on msvcr120.dll. If
your system does not have it, install Visual C++ 2013 Redistibutable Package.

Run examples

Checkout the repository, cd into the root folder, then:

  1. npm install

VGG16 examples

  1. cd example
  2. sh retrieve_vgg16_data.sh

Then, run the VGG16 example.

  1. node example_vgg16.js

You should see something similar to following:

  1. ### Result for ../test/data/Light_sussex_hen.jpg
  2. fc6 out: -29.68303871154785 -52.6440544128418 0.9215406179428101 21.43817710876465 -6.305706977844238 ...
  3. Top 5 categories are:
  4. [8] 0.8902806639671326 n01514859 hen
  5. [86] 0.037541598081588745 n01807496 partridge
  6. [7] 0.03157550096511841 n01514668 cock
  7. [82] 0.017570357769727707 n01797886 ruffed grouse, partridge, Bonasa umbellus
  8. [83] 0.002043411135673523 n01798484 prairie chicken, prairie grouse, prairie fowl
  9. ### Result for ../test/data/honda_nsx.jpg
  10. fc6 out: 14.704771041870117 -10.323609352111816 -32.17032241821289 -9.661919593811035 -14.448777198791504 ...
  11. Top 5 categories are:
  12. [751] 0.6547003388404846 n04037443 racer, race car, racing car
  13. [817] 0.28364330530166626 n04285008 sports car, sport car
  14. [573] 0.02763519063591957 n03444034 go-kart
  15. [511] 0.01738707721233368 n03100240 convertible
  16. [814] 0.004731603432446718 n04273569 speedboat

MNIST examples

In the example folder…

  1. $ node example_mnist.js
  2. ### Result for ../test/data/mnist/0.png
  3. [0] 9792.962890625 Zero
  4. ### Result for ../test/data/mnist/1.png
  5. [1] 4203.07470703125 One
  6. ### Result for ../test/data/mnist/2.png
  7. [2] 7281.75341796875 Two
  8. ### Result for ../test/data/mnist/3.png
  9. [3] 7360.65625 Three
  10. ### Result for ../test/data/mnist/4.png
  11. [4] 3837.8447265625 Four
  12. ### Result for ../test/data/mnist/5.png
  13. [5] 5259.931640625 Five
  14. ### Result for ../test/data/mnist/6.png
  15. [6] 3743.64306640625 Six
  16. ### Result for ../test/data/mnist/7.png
  17. [7] 4321.0859375 Seven
  18. ### Result for ../test/data/mnist/8.png
  19. [8] 3331.339111328125 Eight
  20. ### Result for ../test/data/mnist/9.png
  21. [9] 1424.4774169921875 Nine

Read the comments in the examples for more details.

API

  1. const menoh = require('menoh');

Module methods

menoh.getNativeVersion() => {string}

Returns the version of underlying native menoh (core) library.

menoh.create(onnx_file_path{string}, [cb]) => {Promise}

Returns promise if cb is not provided. The promise resolves to a new instance of ModelBuilder.

ModelBuilder methods

builder.addInput(input_var_name{string}, dims{array}) => {void}

Add an input profile for the given name.

Data type is implicitly set to float32.

builder.addOutput(output_var_name{string}) => {void}

Add an output profile for the given name.

It currently takes no argument other than the name.
Data type is implicitly set to float32.

builder.buildModel(config{object}) => {Model}

Returns an executable model.
The config object can have two properties:

  • backendName {string}: defaults to “mkldnn” or explicitly set it to “mkldnn” always.
  • backendConfig {string}: a JSON string. defaults to “” or set to “” always.

You may build more than one model from the same builder.

Model methods

model.getProfile(var_name{string}) => {object}

Returns a profile information for the given name.
The returned object has following properties:

  • dims {array}: Dimensions of the attached buffer. (e.g. [1, 3, 244, 244])
  • buf {Buffer}: Reference to the buffer attached to the variable.
  • dtype {string}: Data type.

Current revision supports only one data type, “float32”.

model.run(cb) => {Promise}

Run inference. It returns promise if cb is not provided. The actual inference takes place
in a background worker thread. You may run a different models concurrently to take advantage of
available CPU cores.

model.setInputData(input_var_name{string}, data{array})

DEPREACATED. Use model.getProfile() instead.

Sets input data for the give input name.

model.getOutput(output_var_name) => {object}

DEPREACATED. Use model.getProfile() instead.

Returns output object generated during model.run() for the given output name.
The output object has following properties:

  • dims {array}: Output data dimensions. (e.g. [1, 3, 244, 244])
  • data {array}: Output data (flat array).

Limitations

  • You may not call run() on the same model more than once concurrently. The second run() will
    fail with an error. Consider building another model for the concurrent operations.