From Keras to SOFIE

SOFIE is a system developed in ROOT that automatically creates a fast c++ code for inference of machine learning models in ONNX format, which is used when such models are trained with ROOT. However, the format requirement raises a problem when external libraries in python, such as keras and pytorch, are used to train the models, as they use a different format. To enable usage of SOFIE with neural networks trained with the popular keras library in python, we created an automatic parser from keras models into an intermediate class used by SOFIE, thus enabling the usage of advanced neural networks - CNN, RNN, etc. - with the SOFIE system. The parser is reliable and safe to use, and will be incorporated in the ROOT library.