m-badri / fortran-tf-lib

A library for directly calling TensorFlow / Keras ML models from Fortran.

Fortran-TF-lib

Code and examples for directly calling Tensorflow ML models from Fortran.
For calling PyTorch from Fortran see the FTorch repository.

Contents

• Description
• Installation
• Usage
• Examples
• License
• Contributions
• Authors and Acknowledgment
• Users

Description

It is desirable be able to run machine learning (ML) models directly in Fortran. Such models are often trained in some other language (say Python) using popular frameworks (say TensorFlow) and saved. We want to run inference on this model without having to call a Python executable. To achieve this we use the existing TensorFlow C interface.

This project provides a library enabling a user to directly couple their TensorFlow models to Fortran code. We provide installation instructions for the library as well as instructions and examples for performing coupling. This library implements only enough of the TensorFlow C API to allow inference, no training.

Project status: This project is currently in pre-release with documentation and code being prepared for a first release. As such breaking changes may be made. If you are interested in using this library please get in touch.

Installation

Dependencies

To install the library requires the following to be installed on the system:

• cmake >= 3.1
• TensorFlow C API, download from https://www.tensorflow.org/install/lang_c¹
• Fortran and C compilers

¹ Note that this page sometimes does not list the latest version of the library. You can try altering the library download URLs on the page to reflect the newest version. E.g. if the URL ends ...-2.11.tar.gz try changing it to ...-2.13.tar.gz.

Library installation

To build and install the library:

1. Navigate to the location in which you wish to install the source and run:

   git clone [email protected]:Cambridge-ICCS/fortran-tf-lib.git
   

   to clone via ssh, or

   git clone https://github.com/Cambridge-ICCS/fortran-tf-lib.git
   

   to clone via https.

2. Navigate into the library directory by running:

   cd fortran-tf-lib/fortran-tf-lib/
   

3. Create a build directory and execute cmake from within it using the relevant flags:

   mkdir build
   cd build
   cmake .. -DCMAKE_BUILD_TYPE=Release
   

   It is likely that you will need to provide at least the TENSORFLOW_LOCATION flag.
   The Fortran compiler must be the same one that you are planning to compile your Fortran code with. It is advisable to use C and Fortran compilers from the same provider.

   The following CMake flags are available and can be passed as arguments through -D<Option>=<Value>:

   Option	Value	Description
   CMAKE_Fortran_COMPILER	ifort / gfortran	Specify a Fortran compiler to build the library with. This should match the Fortran compiler you're using to build the code you are calling this library from.
   CMAKE_C_COMPILER	icc / gcc	Specify a C compiler to build the library with.
   TENSORFLOW_LOCATION2	</path/to/tensorflow/>	Location of TensorFlow C API installation1.
   CMAKE_INSTALL_PREFIX	</path/to/install/lib/at/>	Location at which the library files should be installed. By default this is /usr/local.
   CMAKE_BUILD_TYPE	Release / Debug	Specifies build type. The default is Debug, use Release for production code.

   ² This should be the absolute path to where the TensorFlow C API mentioned in step 1 has been installed. CMake will look in TENSORFLOW_LOCATION and TENSORFLOW_LOCATION/lib for the TensorFlow library libtensorflow.so.

4. Make and install the code to the chosen location with:

   make
   make install
   

   This will place the following directories at the install location:

   • CMAKE_INSTALL_PREFIX/include/ - contains mod files
   • CMAKE_INSTALL_PREFIX/lib64/ - contains cmake directory and .so files

Usage

In order to use fortran-tf users will typically need to follow these steps:

1. Save a TensorFlow model in the Keras SavedModel format.
2. Write Fortran using the fortran-tf-lib bindings to use the model from within Fortran.
3. Build and compile the code, linking against fortran-tf-lib.

1. Saving the model

The trained model needs to be exported. This can be done from within your code using the model.save functionality from within python. Note that the TensorFlow C API currently (version 2.13) only supports the Keras "v2" format so you must specify format='tf':

import tensorflow as tf
# construct model (e.g. model=tf.keras.Model(inputs, outputs))
# or load one (e.g. model=tf.keras.models.load_model('/path/to/model'))
model.save("my_model", format='tf')

2. Using the model from Fortran

To use the trained TensorFlow model from within Fortran we need to import the TF_Interface module and use the binding routines to load the model, construct the tensors, and run inference.

A very simple example is given below. For more detailed documentation please consult the API documentation, source code, and examples.

This minimal snippet loads a saved TensorFlow model, creates an input consisting of a 1x32 matrix (with arbitrary values), and runs the model to infer the output. If you use the model provided in the test case this code will produce the indicated output value.

program test_program
use TF_Types
use TF_Interface
use iso_c_binding
implicit none

type(TF_Session) :: session
type(TF_SessionOptions) :: sessionoptions
type(TF_Graph) :: graph
type(TF_Status) :: stat
type(TF_Output), dimension(1) :: input_tfoutput, output_tfoutput
character(100) :: vers
character(100), dimension(1) :: tags
type(TF_Tensor), dimension(1) :: input_tensors, output_tensors, test_tensor
type(TF_Operation), dimension(1) :: target_opers

real, dimension(32), target :: raw_data
real, dimension(:), pointer :: output_data_ptr
integer(kind=c_int64_t), dimension(2) :: input_dims
integer(kind=c_int64_t), dimension(2) :: output_dims
type(c_ptr) :: raw_data_ptr
type(c_ptr) :: output_c_data_ptr

raw_data = (/ &
        0.71332126, 0.81275973, 0.66596436, 0.79570779, 0.83973302, 0.76604397, &
        0.84371391, 0.92582056, 0.32038017, 0.0732005, 0.80589203, 0.75226581, &
        0.81602784, 0.59698078, 0.32991729, 0.43125108, 0.4368422, 0.88550326, &
        0.7131253, 0.14951148, 0.22084413, 0.70801317, 0.69433906, 0.62496564, &
        0.50744999, 0.94047845, 0.18191579, 0.2599102, 0.53161889, 0.57402205, &
        0.50751284, 0.65207096 &
        /)


input_dims = (/ 1, 32 /)
output_dims = (/ 1, 1 /)
tags(1) = 'serve'

! Print TensorFlow library version
call TF_Version(vers)
write(*,*)'Tensorflow version', vers

sessionoptions = TF_NewSessionOptions()
graph = TF_NewGraph()
stat = TF_NewStatus()

! Load session (also populates graph)
session = TF_LoadSessionFromSavedModel(sessionoptions, '/path/to/model', tags, 1, &
    graph, stat)

if (TF_GetCode( stat ) .ne. TF_OK) then
    call TF_Message( stat, vers )
    write(*,*)'woops', TF_GetCode( stat ), vers
    call abort
endif

call TF_DeleteSessionOptions(sessionoptions)

input_tfoutput(1)%oper = TF_GraphOperationByName( graph, "serving_default_input_1" )
input_tfoutput(1)%index = 0
if (.not.c_associated(input_tfoutput(1)%oper%p)) then
    write(*,*)'input not associated'
    stop
endif

output_tfoutput(1)%oper = TF_GraphOperationByName( graph, "StatefulPartitionedCall" )
output_tfoutput(1)%index = 0
if (.not.c_associated(output_tfoutput(1)%oper%p)) then
    write(*,*)'output not associated'
    stop
endif

! Bind the input tensor
raw_data_ptr = c_loc(raw_data)
input_tensors(1) = TF_NewTensor( TF_FLOAT, input_dims, 2, raw_data_ptr, int(128, kind=c_size_t) )

! Run inference
call TF_SessionRun( session, input_tfoutput, input_tensors, 1, output_tfoutput, output_tensors, 1, &
    target_opers, 0, stat )
if (TF_GetCode( stat ) .ne. TF_OK) then
    call TF_Message( stat, vers )
    write(*,*) TF_GetCode( stat ), vers
    call abort
endif

! Bind output tensor
call c_f_pointer( TF_TensorData( output_tensors(1)), output_data_ptr, shape(output_data_ptr) )
write(*,*)'output data', output_data_ptr(1)

if ((output_data_ptr(1) - -0.479371) .gt. 1e-6) then
    write(*,*)'Output does not match, FAILED!'
else
    write(*,*)'Output is correct, SUCCESS!'
endif


! Clean up
call TF_DeleteTensor( input_tensors(1) )
call TF_DeleteTensor( output_tensors(1) )
call TF_DeleteGraph( graph )
call TF_DeleteSession( session, stat )
call TF_DeleteStatus( stat )

end program test_program

Generating code with process_model

The example code above illustrates a problem with the TensorFlow C API that our Fortran wrapper cannot fix. To load a model, the library requires that the caller knows certain rather opaque model parameters beforehand. Often, the values in the example above will work for the tags parameter to TF_LoadSessionFromSavedModel. However, the values needed for TF_GraphOperationByName (in this case serving_default_input_1, etc) are more likely to be different.

To address this, we provide a Python script, process_model that will read a Keras SavedModel and output a simple Fortran module intended to provide a base for the user to start from. The appropriate values will be read from the model and hard-coded into the Fortran code.

E.g.

process_model -o fortran_code.f90  my_model

3. Build the code

The code now needs to be compiled and linked against our installed library.

CMake

If our project were using cmake we would need the following in the CMakeLists.txt file to find the the tf-lib installation and link it to the executable.

This can be done by adding the following to the CMakeLists.txt file:

find_package(FortranTensorFlow)
target_link_libraries( <executable> PRIVATE FortranTensorFlow::fortran-tf )
message(STATUS "Building with Fortran TensorFlow coupling")

and using the -DCMAKE_PREFIX=</path/to/fortran-tf-libs/lib64/cmake> flag when running cmake.

When running the generated code you may also need to add the location of the .so files to your LD_LIBRARY_PATH unless installing in a default location:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<path/to/fortran-tf-libs>/lib64

Examples

Examples of how to use this library will be provided in the examples directory.
They demonstrate different functionalities and are provided with instructions to modify, build, and run as necessary.

License

Copyright © ICCS

Fortran-TF-Lib is distributed under the MIT Licence.

Contributions

Contributions and collaborations are welcome.

For bugs, feature requests, and clear suggestions for improvement please open an issue.

If you have built something upon Fortran-TF-Lib that would be useful to others, or can address an open issue, please fork the repository and open a pull request.

Code of Conduct

Everyone participating in the Fortran-TF-Lib project, and in particular in the issue tracker, pull requests, and social media activity, is expected to treat other people with respect and, more generally, to follow the guidelines articulated in the Python Community Code of Conduct.

Authors and Acknowledgment

Fortran-TF-Lib is written and maintained by the ICCS

Notable contributors to this project are:

• @SimonClifford

See Contributors for a full list.

Used by

The following projects make use of this code or derivatives in some way:

• DataWave - MiMA ML

Are we missing anyone? Let us know.