BMXNet 2 // Hasso Plattner Institute

A fork of the deep learning framework mxnet to study and implement quantization and binarization in neural networks.

This project is based on the first version of BMXNet, but is different in that it reuses more of the original MXNet operators. This aim was to have only minimal changes to C++ code to get better maintainability with future versions of mxnet.

mxnet version

This version of BMXNet 2 is based on: mxnet v1.4.1

News

See all BMXNet changes: Changelog.

• May 21, 2019
  • Model converter for deployment released (Test & Example)
• Sep 01, 2018
  • We rebuilt BMXNet to utilize the new Gluon API for better maintainability
  • To build binary neural networks, you can use drop in replacements of convolution and dense layers (see Usage):
  • Note that this project is still in beta and changes might be frequent

Setup

We use CMake to build the project. Make sure to install all the dependencies described here. If you install CUDA 10, you will need CMake >=3.12.2

Adjust settings in cmake (build-type Release or Debug, configure CUDA, OpenBLAS or Atlas, OpenCV, OpenMP etc.).

Further, we recommend Ninja as a build system for faster builds (Ubuntu: sudo apt-get install ninja-build).

git clone --recursive https://github.com/hpi-xnor/BMXNet-v2.git # remember to include the --recursive
cd BMXNet-v2
mkdir build && cd build
cmake .. -G Ninja # if any error occurs, apply ccmake or cmake-gui to adjust the cmake config.
ccmake . # or GUI cmake
ninja

Build the MXNet Python binding

Step 1 Install prerequisites - python, setup-tools, python-pip and numpy.

sudo apt-get install -y python-dev python3-dev virtualenv
wget -nv https://bootstrap.pypa.io/get-pip.py
python3 get-pip.py
python2 get-pip.py

Step 1b (Optional) Create or activate a virtualenv.

Step 2 Install the MXNet Python binding.

cd <mxnet-root>/python
pip install -e .

If your mxnet python binding still not works, you can add the location of the libray to your LD_LIBRARY_PATH as well as the mxnet python folder to your PYTHONPATH:

$ export LD_LIBRARY_PATH=<mxnet-root>/build/Release
$ export PYTHONPATH=<mxnet-root>/python

Training

Make sure that you have a new version of our example submodule example/bmxnet-examples:

cd example/bmxnet-examples
git checkout master
git pull

Examples for hyperparameters are documented in the Wiki.

Inference

To speed up inference and compress your model, you need to save it as a symbol (not with gluon) and afterwards convert it with the model-converter. Please check the corresponding test case.

build/tools/binary_converter/model-converter model-0000.params

Tests

To run BMXNet specific tests install pytest:

pip install pytest

Then simply run:

pytest tests/binary

Usage

We added binary versions of the following layers of the gluon API:

• gluon.nn.Dense -> gluon.nn.QDense
• gluon.nn.Conv1D -> gluon.nn.QConv1D
• gluon.nn.Conv2D -> gluon.nn.QConv2D
• gluon.nn.Conv3D -> gluon.nn.QConv3D

Overview of Changes

We added three functions det_sign (ada4ea1d), round_ste (044f81f0) and contrib.gradcancel to MXNet (see src/operator/contrib/gradient_cancel[-inl.h|.cc|.cu]).

The rest of our code resides in the following folders/files:

• Examples are in a submodule in example/bmxnet-examples
• Tests are in tests/binary
• Layers are in python/mxnet/gluon/nn/binary_layers.py
• Converter is in tools/binary_converter

For more details see the Changelog.

Citing BMXNet 2

Please cite our paper about BMXNet 2 in your publications if it helps your research work:

@article{bmxnetv2,
  title = {Training Competitive Binary Neural Networks from Scratch},
  author = {Joseph Bethge and Marvin Bornstein and Adrian Loy and Haojin Yang and Christoph Meinel},
  journal = {ArXiv e-prints},
  archivePrefix = "arXiv",
  eprint = {1812.01965},
  Year = {2018}
}

References

• XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
• Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
• DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients