Page History

In the life of a physicist, the time arrives when you get disgusted by the software dependencies that evaluating a simple neural network requires and you want the plain, pure and clean function to be evaluated in a whatever context. 

For me it was to run a trained algorithm on a Arduino board that we were using in the lab to stabilize the temperature of a device, for a colleague it was to use HLS to push a simple neural network in his FPGA project for another friend it was to feed into the TTree::Draw function of ROOT the evaluator of a neural network trained in keras to avoid rewriting the rest of the software stack originally designed for TMVA. 

In this tutorial we propose an example of how to translate a simple neural network into single header file (.h) that can be exported and #included a whatever C application without a single dependency to get the trained algorithm hard-coded and compiled together with your program. 

Colleagues from particle physics have used a similar approach to provide a tensorflow-free python evaluator of (more complicated) neural networks. Check out the tf.deploy package for additional information: https://github.com/riga/tfdeploy

Back to C/C++, you can checkout the LWTNN project born within the ATLAS community (https://github.com/lwtnn/lwtnn), which needs few external dependencies and cannot be used on Arduino for instance, but allows to load dynamically the neural networks from disk avoiding all that horrible hard-coding. 

But since this is a tutorial and we are not writing production code, let's have a look. 

We will take as problem the MNIST problem which is a kind of Hello World for multi-category classification tasks in machine learning, in a way that we can also review this exercise that cannot be missing in a ML-related web page. 

The other reason is that the original data that we used to train this are not public. 

Static jupter notebook: https://github.com/landerlini/MLINFN-TutorialNotebooks/blob/master/TranscribeAnMLPinC.ipynb

Requirements

To run this exercise you will need a plain installation of python with tensorflow 1.x (should work with tensorflow 2.x, but it was not tested).

Contents

1. What is multi-category classification
2. One-hot encoding
3. Training with high-level Keras APIs
4. C code generation for the algorithm (a bit weird)
5. C code generation for the input dataset (very weird) 
6. Test of the performance of the converted algorithm.