README in progress, for now please refer to

Child Growth Monitor Machine Learning

Introduction

This project uses machine learning to identify malnutrition from 3D scans of children under 5 years of age. This one-minute video explains.

Getting started

As a first step, please make sure that you work with all the notebooks provided in the skel-folder. This will give you a quick and thourough introduction to the current state-of-the-implementation.

Requirements

Training the models realistically requires using GPU computation. A separate backend project is currently developing the DevOps infrastructure to simplify this.

You will need:

Python 3
TensorFlow GPU
Keras

Create documentation

Do this:

cd docs
make html

Then open docs/build/html/index.html

Installation of cgmcore module.

You can install the utilities as a module:

pip install git+https://github.com/Welthungerhilfe/cgm-ml.git

Installation of Point Cloud Library

Steps to install pcl-1.9.0 on the virtual machines used during the datathon. Unfortunately, these steps still don't make it easy to use python-pcl due to pcl-1.7 already being present on the machines.

conda deactivate # get out of datathon env
conda deactivate # get out of py35 env
sudo apt -y install libflann1.8 libboost1.58-all-dev libeigen3-dev libproj-dev
wget https://github.com/PointCloudLibrary/pcl/archive/pcl-1.9.0.tar.gz
tar xvf pcl-1.9.0.tar.gz
cd pcl-pcl-1.9.0/ && mkdir build && cd build
cmake -DCMAKE_CXX_FLAGS="-L/usr/lib -lz" ../
make -j2
sudo make -j2 install
conda activate py35
conda activate datathon

Installation

Linux (Ubuntu)

These steps provide an example installation on a local Ubuntu workstation from scratch:

Install Ubuntu Desktop 18.04.1 LTS
Install NVIDIA drivers
Please note that after rebooting, the secure boot process will prompt you to authorize the driver to use the hardware via a MOK Management screen.

sudo add-apt-repository ppa:graphics-drivers
sudo apt-get update
sudo apt-get install nvidia-390
sudo reboot now

Install Anaconda with Python 3.6

conda update anaconda
conda update python
conda update --all
conda create --name cgm
source activate cgm
conda install tensorflow-gpu
conda install ipykernel
conda install keras
conda install vtk progressbar2 glob2 pandas
pip install --upgrade pip
pip install git+https://github.com/daavoo/pyntcloud

macOS

Tensorflow dropped GPU support on macOS. Otherwise the installation is similar to the one on Linux above.

Install Anaconda with Python 3.6

conda update anaconda
conda update python
conda update --all
conda create --name cgm
source activate cgm
conda install tensorflow
conda install ipykernel
conda install keras
conda install vtk progressbar2 glob2 pandas
pip install --upgrade pip
pip install git+https://github.com/daavoo/pyntcloud

Known issues on macOS

Saving prepared datasets > 2GB fails
- Error: OSError: [Errno 22] Invalid argument
- Our current datasets are > 2GB
- We save them as Pickle files as preparation for the training
- Python bug ticket: https://bugs.python.org/issue24658
- Workaround 1: Reduce dataset size
- Workaround 2: Apply https://stackoverflow.com/a/38003910

Experimental GPU support on macOS

Install tensorflow-gpu 1.1 via pip: https://www.tensorflow.org/versions/r1.1/install/install_mac.

Dataset access

Data access is provided on as-needed basis following signature of the Welthungerhilfe Data Privacy & Commitment to Maintain Data Secrecy Agreement. If you need data access (e.g. to train your machine learning models), please contact Markus Matiaschek for details.

Data preprocessing

Read the README at dataset-database.

Contributing

Please see CONTRIBUTING.md for details.

Versioning

Our releases use semantic versioning. You can find a chronologically ordered list of notable changes in CHANGELOG.md.

License

This project is licensed under the GNU General Public License v3.0. See LICENSE for details and refer to NOTICE for additional licensing notes and use of third-party components.

christianpfitzner/cgm-ml