is there any way to convert NumSharp.NDArray to Numpy.NDarray? #383

Activity

1. Nucs commented on Jan 6, 2020

   

   Nucs

   on Jan 6, 2020

   Member

   More actions

   Me and @henon have collaborated just two days ago on a solution for that,
   The trick is to create a python scope, load a script that takes in a pointer-length and returns an np.ndarray
   Remember! you have to keep reference to the original NDArray, otherwise the NDArray will be disposed and will release the memory. (@henon is there a tag field we can use in NDarray?)

   Note: the code below is psuedo, the rest of the code can be found here: ConsoleApp9.zip

   numpy_interop.py as embedded resource

   import numpy as np
   import ctypes
   
   def to_numpy(ptr, bytes, dtype, shape):
       return np.frombuffer((ctypes.c_uint8*(bytes)).from_address(ptr), dtype).reshape(shape)

   PythonExtensions.cs

   public static class PythonExtensions {
   
       ...
       private static PyScope NumpyInteropScope;
       private static PyObject NumpyConverter;
   
       public static unsafe NDarray ToNumpyNET(this NDArray nd) {
           using (Py.GIL()) {
               if (NumpyInteropScope == null) {
                   NumpyInteropScope = Py.CreateScope();
                   NumpyInteropScope.ExecResource("numpy_interop.py");
                   NumpyConverter = NumpyInteropScope.GetFunction("to_numpy");
               }
   
               return new NDarray(NumpyConverter.Invoke(new PyLong((long) nd.Unsafe.Address),
                   new PyLong(nd.size * nd.dtypesize),
                   new PyString(nd.dtype.Name.ToLowerInvariant()),
                   new PyTuple(nd.shape.Select(dim => (PyObject) new PyLong(dim)).ToArray())));
           }
       }
   }

   main.cs

   using numsharp_np = NumSharp.np;
   using numpy_np = Numpy.np;
   static void Main(string[] args) {
       var numsharp_nd = numsharp_np.arange(10);
       var numpy_nd = numsharp_nd.ToNumpyNET();
       numpy_nd[1] = (NDarray) 5;
       Debug.Assert(numsharp_nd[1].array_equal(5));
       Console.WriteLine("numsharp_nd[1].array_equal(5) is indeed true!");
   }
2. Nucs commented on Jan 6, 2020

   

   Nucs

   on Jan 6, 2020

   Member

   More actions

   Added some changes that will make sure the NDarray will hold reference to the NDArray as long as the PyObject doesnt change internally.

   ConsoleApp9.zip
3. henon commented on Jan 6, 2020

   

   henon

   on Jan 6, 2020

   Contributor

   More actions

   Remember! you have to keep reference to the original NDArray, otherwise the NDArray will be disposed and will release the memory. (@henon is there a tag field we can use in NDarray?)

   Not yet, but I will add one in Numpy.NET and support this.

   🚀React with 🚀1Reacted by Eli Belash
4. davidvct commented on Feb 24, 2023

   

   davidvct

   on Feb 24, 2023

   More actions

   Hi, could you provide instruction on how to use the ConsoleApp9.zip? And does it able to convert Numpy array back to NumSharp array?
5. henon commented on Feb 24, 2023

   

   henon

   on Feb 24, 2023

   Contributor

   More actions

   @davidvct, if I remember correctly I have added support for copying data to and from Numpy.NET like @Nucs did in this solution. Read this to see how: https://github.com/SciSharp/Numpy.NET#create-a-numpy-array-from-a-c-array-and-vice-versa

   👍React with 👍2Reacted by Eli Belash and davidvct
6. Nucs commented on Sep 7, 2026

   

   Nucs

   on Sep 7, 2026

   Last edited by Nucs

   Member

   More actions

   @lelelemonade @davidvct, the manual pointer-scope trick from this thread is superseded. The new NumSharp.Interop.pythonnet package (0.70.0, #628) crosses both ways zero-copy, with GC-safe leases (no more "keep a reference or it gets disposed").

   Set up the Python engine once at startup - either point pythonnet at an installed Python, or deploy one from NuGet with Python.Included (no system install needed):

   using NumSharp;
   using NumSharp.Interop.PythonNet;
   using Python.Runtime;
   
   // Option A - existing Python: the PYTHONNET_PYDLL env var, or the DLL directly
   Runtime.PythonDLL = "python312.dll";   // or Environment.GetEnvironmentVariable("PYTHONNET_PYDLL")
   PythonEngine.Initialize();
   
   // Option B - deploy an embedded Python from NuGet (dotnet add package Python.Included)
   await Python.Included.Installer.SetupPython();
   await Python.Included.Installer.PipInstallModule("numpy");
   PythonEngine.Initialize();

   Explicit:

   var nd = np.arange(6).reshape(2, 3);
   
   using (Py.GIL())
   {
       // zero-copy: numpy shares NumSharp's buffer
       dynamic npArr = nd.ToNumpy();
       Console.WriteLine(npArr.sum());   // 15
   }

   Implicit (register the codec once, then NumSharp arrays auto-convert):

   NDArrayPythonInterop.RegisterCodec();   // once per process
   
   var nd = np.arange(6).reshape(2, 3);
   using (Py.GIL())
   {
       dynamic numpy = Py.Import("numpy");
       Console.WriteLine(numpy.sum(nd));   // 15   (nd auto-encoded by the codec)
   }

   Requires NumSharp.Interop.pythonnet (dotnet add package NumSharp.Interop.pythonnet). Closing.
7. Nucs

   closed this as completedon Sep 7, 2026