process_model

The process_model tool reads a TensorFlow SavedModel and outputs Fortran code to interface it to the fortran-tf-lib

Installing

In a suitable Python environment do:

pip install git+https://github.com/Cambridge-ICCS/process_model.git

Note that as of 20/01/23 there is no tensorflow package in Pypi for Python >= 3.11.

Running the tool

The pip install will place a process_model command in the PATH. To use it, run it against one or more TensorFlow SavedModel models.

process_model model_1 model_2 ...

The tool will output Fortran code to standard output, or to the file specified with the -o option.

Using the resulting Fortran

The output is a module, named ml_module by default. It has procedures called ml_module_init, ml_module_calc, ml_module_finish. It also may have some *_associate_tensor routines tailored for the inputs of the model. So if the model expects a Tensor of type TF_FLOAT and of shape [-1, 40] then there will be a r32_2_associate_tensor routine to generate appropriately shaped and typed tensors from Fortran arrays.

The ml_module_init routine should be called once, before using calc. It loads the models into module variables.

Worked example

API reference