Menoh: fast DNN inference library with multiple programming language support
Clone or download
msakai Merge pull request #173 from pfnet-research/travis-no-uninstall-oclint
stop uninstalling oclint as Travis-CI itself removes it automatically
Latest commit 65d9e17 Nov 13, 2018

README.md

Menoh

travis Build status Coverity Scan Build Status

Menoh is DNN inference library with C API.

Menoh is released under MIT License.

DISCLAIMER: Menoh is still experimental. Use it at your own risk. In particular not all operators in ONNX are supported, so please check whether the operators used in your model are supported. We have checked that VGG16 and ResNet50 models converted by onnx-chainer work fine.

Document

This codebase contains C API and C++ API.

Goal

  • DNN Inference with CPU
  • ONNX support
  • Easy to use.

Related Projects

Installation using package manager or binary packages

  • For Windows users, prebuild libraries are available (see release) and Nuget package is available.
  • For macOS user, Homebrew tap repository is available.
  • For Ubuntu user, binary packages are available.
    $ curl -LO https://github.com/pfnet-research/menoh/releases/download/v1.1.1/ubuntu1604_mkl-dnn_0.16-1_amd64.deb
    $ curl -LO https://github.com/pfnet-research/menoh/releases/download/v1.1.1/ubuntu1604_menoh_1.1.1-1_amd64.deb
    $ curl -LO https://github.com/pfnet-research/menoh/releases/download/v1.1.1/ubuntu1604_menoh-dev_1.1.1-1_amd64.deb
    $ sudo apt install ./ubuntu1604_*_amd64.deb
    
    If you are using Ubuntu 18.04, please replace 1604 with 1804.

Installation from source

Requirements

  • MKL-DNN Library (0.14 or later)
  • Protocol Buffers (2.6.1 or later)

Build

Execute following commands in root directory.

python scripts/retrieve_data.py
mkdir build && cd build
cmake ..
make

See BUILDING.md for details.

Installation

Execute following command in build directory created at Build section.

make install

Run VGG16 example (it can run ResNet-50 as well)

Execute following command in root directory.

./example/vgg16_example_in_cpp

Result is here

vgg16 example
-18.1883 -26.5022 -20.0474 13.5325 -0.107129 0.76102 -23.9688 -24.218 -21.6314 14.2164 
top 5 categories are
8 0.885836 n01514859 hen
7 0.104591 n01514668 cock
86 0.00313584 n01807496 partridge
82 0.000934658 n01797886 ruffed grouse, partridge, Bonasa umbellus
97 0.000839487 n01847000 drake

You can also run ResNet-50

./example/vgg16_example_in_cpp -m ../data/resnet50.onnx

Please give --help option for details

./example/vgg16_example_in_cpp --help

Run test

Setup chainer

Then, execute following commands in root directory.

python scripts/gen_test_data.py
cd build
cmake -DENABLE_TEST=ON ..
make
./test/menoh_test.out

Current supported operators

Activation functions

  • Elu
  • LeakyRelu
  • Relu
  • Softmax
  • Tanh

Array manipulations

  • Concat

Neural network connections

  • Conv
  • ConvTranspose
  • FC

Mathematical functions

  • Abs
  • Add
  • Sqrt
  • Sum

Normalization functions

  • BatchNormalization
  • LRN

Spatial pooling

  • AveragePool
  • GlobalAveragePool
  • GlobalMaxPool
  • MaxPool

License

Menoh is released under MIT License. Please see the LICENSE file for details.

Pre-trained models downloaded via retrieve_data.py were converted by onnx-chainer. The original models were downloaded via ChainerCV. Check scripts/generate_vgg16_onnx.py and scripts/generate_resnet50_onnx.py and see the LICENSE of ChainerCV about each terms of use of the pre-trained models.