Extending and Developing the Medial C++ API for Python

This document outlines the low-level C++ details required to further develop the Python bindings for our API.

Objectives

The goals are to:

1. Use our API from Python.
2. Employ Python as a "Glue Language".
3. Enable prototyping, exploration, and discovery in Python.
4. Ensure interoperability with other frameworks.
5. Minimize maintenance overhead when updating code.

Implementation Approach

For maintainability, SWIG is chosen as the binding solution. Other options considered:

• Cython: Adds another language to the process.
• Boost::python: Limited support/resources; requires special definitions for each exported member.
• ctypes: Direct C calls, but parameter definitions are error-prone and may lack binary compatibility across platforms.

Why SWIG?

• Mature project with extensive resources. TensorFlow uses SWIG for example.
• Handles memory and ownership complexities.
• Good NumPy support.

Our Implementation: "MedPyExport"

• Located in $MR_ROOT/Libs/Internal/MedPyExport.
• Wraps classes and exports only necessary functions for Python.
• Minimal learning curve for contributors familiar with C++.
• SWIG interface files (.i) are auto-generated during build.

Notes & Pitfalls

• Always use std::vector for vectors.
• Restart Jupyter Notebook after adding new functions/classes to reload wrappers.
• Use distinct names for NumPy parameters of different types.
• Method overloading is not supported (may be possible to implement).
• Avoid "big-data" loops; use NumPy/Pandas for vectorized operations.
• Build system is separate from the main CMake files.
• No direct Pandas DataFrame conversion yet.
• SWIG scans only simple C++ headers; keep implementation in .cpp files.
• Compilation works on both Windows and Linux (tested execution on Linux/Jupyter).

Build System

• Targeted for Linux machines
• Windows compilation works but does not produce a Python-loadable binary.
• Uses a specialized CMake file for SWIG integration.

Directory Structure

Located in $MR_ROOT/Libs/Internal/MedPyExport:

• MedPyExport: Source files.
• generate_binding: CMake and binding generation files.
• MedPython: SWIG interface files.
  • scripts: Helper scripts for compilation.

Key Files

• make.sh: Bash script to start CMake build.
• MedPython/MedPython.i: Main SWIG interface.
• MedPython/medial-numpy.i: NumPy SWIG interface.
• MedPython/MedPython.h: Entry point for headers scanned by SWIG.
• MedPython/MedPython.c: Entry point for C files scanned by SWIG.
• MedPython/pythoncode.i: Python code for the binding.
• MedPython/scripts/make_apply.py: Script to generate NumPy apply directives.
• MedPython/apply_directives.i: Auto-generated SWIG directives.

Extending the Code

Example: Exporting a Class

// .h file for the class 'PidRepository'
#include "MedPyCommon.h"
class MedPidRepository;
class MPPidRepository {
public:
    MedPidRepository* o;
    // ...
};

• Include MedPyCommon.h for utilities/macros.
• Exported class names start with MP to avoid conflicts.
• Use a pointer to the wrapped object.
• Keep headers simple for SWIG; implementation in .cpp files.

Add all exported headers to MedPyExport.h:

#ifndef __MED_PY_EXPORT_H
#define __MED_PY_EXPORT_H
#include "MedPyExportExample.h"
#include "MPPidRepository.h"
#include "MPDictionary.h"
// ...
#endif

Python Usage:

import medpython as med
rep = med.PidRepository()

Adding a New Class

class MPPidRepository {
public:
    MedPidRepository* o;
    MPPidRepository();
    ~MPPidRepository();
    int read_all(const string &conf_fname);
    string dict_name(int section_id, int id);
    std::vector<bool> dict_prep_sets_lookup_table(int section_id, const std::vector<std::string> &set_names);
    // ...
};

• Basic types are mapped automatically.
• Use std::vector for vectors.

Implementation Example:

#include "MPPidRepository.h"
#include "InfraMed/InfraMed/MedPidRepository.h"
MPPidRepository::MPPidRepository() : o(new MedPidRepository()) {}
MPPidRepository::~MPPidRepository() { delete o; }
int MPPidRepository::read_all(const std::string &conf_fname) { return o->read_all(conf_fname); }
string MPPidRepository::dict_name(int section_id, int id) { return o->dict.name(section_id, id); }

Properties

Implement getter/setter methods with MEDPY_GET_ and MEDPY_SET_ prefixes:

class MPSamples {
    int MEDPY_GET_time_unit();
    void MEDPY_SET_time_unit(int new_time_unit);
};

Python Usage:

>>> s = med.Samples()
>>> s.time_unit
1

• Omit the setter for read-only properties.

Static Const Variables

Mapped to class variables in Python:

class MPTime {
public:
    static const int Undefined;
    static const int Date;
    static const int Years;
};

Implementation:

const int MPTime::Undefined = MedTime::Undefined;
const int MPTime::Date = MedTime::Date;

Python Usage:

print(med.Time.Date)

Iterators

Array Iterator Example:

class MPSigVectorAdaptor {
public:
    int __len__();
    MPSig __getitem__(int i);
};

• Class name ends with VectorAdaptor.
• Implements __len__ and __getitem__.

Map Iterator Example:

class MPStringFeatureAttrMapAdaptor {
    std::map<std::string, FeatureAttr>* o;
public:
    int __len__();
    MPFeatureAttr __getitem__(std::string key);
    void __setitem__(std::string key, MPFeatureAttr& val);
    std::vector<std::string> keys();
};

• Class name ends with MapAdaptor.
• Implements __len__, __getitem__, and optionally __setitem__, etc.

NumPy Arrays

Input Arrays:

class MPPidRepository {
    int read_all(string conf_fname, MEDPY_NP_INPUT(int* pids_to_take, int num_pids_to_take));
}

In-place Arrays:

void MedPyExportExample::numpy_vec_in_out(MEDPY_NP_INPLACE(double* vec, int m));

Output Arrays:

class MPFeatures {
    void MEDPY_GET_weights(MEDPY_NP_OUTPUT(float** float_out_buf, int* float_out_buf_len));
};

Variant Output Example:

void getitem(string key, MEDPY_NP_VARIANT_OUTPUT(void** var_arr, int* var_arr_sz, int* var_arr_type)) {
    // Implementation sets var_arr, var_arr_sz, var_arr_type based on key
    *var_arr_sz = 0;
    if (key == "i")
    {
        *var_arr = (void*)malloc(sizeof(int) * 10);
        *var_arr_sz = 10;
        *var_arr_type = (int)MED_NPY_TYPES::NPY_INT;
        for (int i = 0; i < 10; i++)
            ((*(int**)var_arr))[i] = i * 5;
    }
    else if (key == "d")
    {
        *var_arr = (void*)malloc(sizeof(double) * 20);
        *var_arr_sz = 20;
        *var_arr_type = (int)MED_NPY_TYPES::NPY_DOUBLE;
        for (int i = 0; i < 20; i++)
            ((*(double**)var_arr))[i] = i * 2.5;
    }
    else if (key == "f")
    {
        *var_arr = (void*)malloc(sizeof(float) * 15);
        *var_arr_sz = 15;
        *var_arr_type = (int)MED_NPY_TYPES::NPY_FLOAT;
        for (int i = 0; i < 15; i++)
            ((*(float**)var_arr))[i] = i * 0.33333f;
    }
    else if (key == "n")
    {
        *var_arr = nullptr;
    }
}

Supported NumPy Types:

NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_OBJECT,
NPY_STRING, NPY_UNICODE, NPY_VOID, NPY_DATETIME, NPY_TIMEDELTA, NPY_HALF,
NPY_NTYPES, NPY_NOTYPE, NPY_CHAR, NPY_USERDEF, NPY_NTYPES_ABI_COMPATIBLE

Helper Functions & Macros

• Docstrings:
  MEDPY_DOC(function_or_class_name, docstring)
• Ignore for SWIG:
  MEDPY_IGNORE(...) or #ifdef SWIG ... #endif
• Buffer to Vector:
  buf_to_vector(int_in_buf, int_in_buf_len, idx);
• Vector to Buffer:
  vector_to_buf(o->weights, float_out_buf, float_out_buf_len);