Numpy.NET is the most complete .NET binding for NumPy, which is a fundamental library for scientific computing, machine learning and AI in Python. Numpy.NET empowers .NET developers with extensive functionality including multi-dimensional arrays and matrices, linear algebra, FFT and many more via a compatible strong typed API. Several other SciSharp projects like Keras.NET and Torch.NET depend on Numpy.NET.
Check out this example which uses numpy
operations to fit a two-layer neural network to random data by manually implementing the forward and backward passes through the network.
Numpy and Intellisense: a developer-friendly combination:
If you want to use Numpy.NET you have two options:
Just reference Numpy.dll via Nuget, set your build configuration to x64 and you are good to go. Thanks to Python.Included it doesn't require a local Python installation or will not clash with existing installations.
In certain use cases you might not want the packaged Python and NumPy packages. In that case you reference Numpy.Bare.dll via Nuget. Depending on the Numpy.Bare nuget version will need Python 3.5, 3.6 or 3.7 and Numpy 1.16 installed for it to work. The first two digits of the Numpy.Bare version indicate which Python version is needed for it to run (i.e. Numpy.Bare v3.6.1.1 needs Python 3.6 installed). If you are getting BadImageFormatException switch between x86 and x64 build settings.
Numpy.NET uses Python for .NET to call into the Python module numpy
. However, this does not mean that it depends on a local Python installation! Numpy.NET.dll uses Python.Included which packages embedded Python 3.7 and automatically deploys it in the user's home directory upon first execution. On subsequent runs, it will find Python already deployed and therefor doesn't install it again. Numpy.NET also packages the NumPy wheel and installs it into the embedded Python installation when not yet installed.
Long story short: as a .NET Developer you don't need to worry about Python at all. You just reference Numpy.NET, use it and it will just work, no matter if you have local Python installations or not.
You might ask how calling into Python affects performance. As always, it depends on your usage. Don't forget that numpy
's number crunching algorithms are written in C
so the thin pythonnet
and Python
layers on top won't have a significant impact if you are working with larger amounts of data.
In my experience, calling numpy
from C# is about 4 times slower than calling it directly in Python
while the execution time of the called operation is of course equal. So if you have an algorithm that needs to call into numpy
in a nested loop, Numpy.NET may not be for you due to the call overhead.
All of numpy
is centered around the ndarray
class which allows you to pass a huge chunk of data into the C
routines and let them execute all kinds of operations on the elements efficiently without the need for looping over the data. So if you are manipulating arrays or matrices with thousands or hundreds of thousands of elements, the call overhead will be negligible.
The most performance sensitive aspect is creating an NDarray
from a C#
array, since the data has to be moved from the CLR
into the Python
interpreter. pythonnet
does not optimize for passing large arrays from C#
to Python
but we still found a way to do that very efficiently. When creating an array with np.array( ... )
we internally use Marshal.Copy
to copy the entire C#
-array's memory into the numpy
-array's storage. And to efficiently retrieve computation results from numpy
there is a method called GetData<T>
which will copy the data back to C#
in the same way:
// create a 2D-shaped NDarray<int> from an int[]
var m = np.array(new int[] {1, 2, 3, 4});
// calculate the cosine of each element
var result = np.cos(m);
// get the floating point data of the result NDarray back to C#
var data = result.GetData<double>(); // double[] { 0.54030231, -0.41614684, -0.9899925 , -0.65364362 }
Python/NumPy doesn't have real multi-threading support. There is no advantage in "simultaneously" executing numpy
functions on multiple threads because pythonnet
requires you to use the Global Interpreter Lock (GIL) to lock access to the Python engine exclusively for only one thread at a time. If you have to call Python from a thread other than the main thread you first must release the main thread's mutex by calling PythonEngine.BeginAllowThreads()
or you'll have a deadlock:
var a = np.arange(1000);
var b = np.arange(1000);
// https://github.com/pythonnet/pythonnet/issues/109
PythonEngine.BeginAllowThreads();
Task.Run(()=> {
// when running on different threads you must lock!
using (Py.GIL())
{
np.matmul(a, b);
}
}).Wait();
Above example only serves as a reference on how to call numpy
from a different thread than the main thread. As said before, having multiple background threads that call into Python doesn't give you multi-core processing because of the requirement to lock the GIL. Not doing so will result in access violation exceptions or deadlocks.
The SciSharp team is also developing a pure C# port of NumPy called NumSharp which is quite popular albeit being not quite complete.
There are a couple of other NumPy ports out there featuring subsets of the original library. The only one that matches Numpy.NET in terms of completeness is the IronPython package numpy
which is out of date though. The SciSharp team is committed to keeping Numpy.NET up to date with the original library and to feature as much of the original functionality as possible.
The vast majority of Numpy.NET's code is generated using CodeMinion by parsing the documentation at docs.scipy.org/doc/numpy/. This allowed us to wrap most of the numpy
-API in just two weeks. The rest of the API can be completed in a few more weeks, especially if there is popular demand.
The following API categories have been generated (if checked off)
- Array creation routines
- Array manipulation routines
- Binary operations
- String operations
- Datetime Support Functions
- Data type routines
- Optionally Scipy-accelerated routines(numpy.dual)
- Floating point error handling
- Discrete Fourier Transform(numpy.fft)
- Financial functions
- Functional programming
- Indexing routines
- Input and output
- Linear algebra(numpy.linalg)
- Logic functions
- Masked array operations
- Mathematical functions
- Matrix library(numpy.matlib)
- Miscellaneous routines
- Padding Arrays
- Polynomials
- Random sampling(numpy.random)
- Set routines
- Sorting, searching, and counting
- Statistics
- Window functions
Over 500 functions of all 1800 have been generated, most of the missing functions are duplicates on Matrix, Chararray, Record etc.
We even generated hundreds of unit tests from all the examples found on the NumPy documentation. Most of them don't compile without fixing them because we did not go so far as to employ a Python-To-C# converter. Instead, the tests are set up with a commented out Python-console log that shows what the expected results are and a commented out block of somewhat C#-ified Python code that doesn't compile without manual editing.
Another reason why this process can not be totally automated is the fact that NumPy obviously changed the way how arrays are printed out on the console after most of the examples where written (they removed extra spaces between the elements). This means that oftentimes the results needs to be reformatted manually for the test to pass.
Getting more unit tests to run is very easy though, and a good portion have already been processed to show that Numpy.NET really works. If you are interested in working on the test suite, please join in and help. You'll learn a lot about NumPy on the way.
Since we have taken great care to make Numpy.NET as similar to NumPy itself, you can, for the most part, rely on the official NumPy manual.
Creating an NDarray
from data is easy. Just pass the C# array into np.array(...)
. You can pass 1D, 2D and 3D C# arrays into it.
// create a 2D NDarray
var m = np.array(new int[,] {{1, 2}, {3, 4}}); // the NDarray represents a 2 by 2 matrix
Another even more efficient way (saves one array copy operation) is to pass the data as a one-dimensional array and just reshape the NDarray.
// create a 2D NDarray
var m = np.array(new int[] {1, 2, 3, 4}).reshape(2,2); // the reshaped NDarray represents a 2 by 2 matrix
As you have seen, apart from language syntax and idioms, usage of Numpy.NET is almost identical to Python. However, due to lack of language support in C#, there are some differences which you should be aware of.
You can access parts of an NDarray using array slicing. C# doesn't support the colon syntax in indexers i.e. a[:, 1]
. However, by allowing to pass a string we circumvented this limitation, i.e. a[":, 1"]
. Only the ...
operator is not yet implemented, as it is not very important.
Some NumPy functions like reshape
allow variable argument lists at the beginning of the parameter list. This of course is not allowed in C#, which supports a variable argument list only as the last parameter. In case of reshape
the solution was to replace the variable argument list that specifies the dimensions of the reshaped array by a Shape
object which takes a variable list of dimensions in its constructor:
var a=np.arange(24);
var b=np.reshape(a, new Shape(2, 3, 4)); // or a.reshape(2, 3, 4)
In other cases the problem was solved by providing overloads without the optional named parameters after the variable argument list.
C# is very strict about parameter default values which must be compile time constants. Also, parameters with default values must be at the end of the parameter list. The former problem is usually solved by having a default value of null
instead of the non-compile-time constant and setting the correct default value inside the method body when called with a value of null
.
Python operators that are not supported in C# are exposed as instance methods of NDarray
:
Python | C# |
---|---|
// | floordiv() |
** | pow() |
In NumPy you can write a*=2
which doubles each element of array a
. C# doesn't have a *=
operator and consequently overloading it in a Python fashion is not possible. This limitation is overcome by exposing inplace operator functions which all start with an 'i'. So duplicating all elements of array a
is written like this in C#: a.imul(2);
. The following table lists all inplace operators and their pendant in C#
Python | C# |
---|---|
+= | iadd() |
-= | isub() |
*= | imul() |
/= | idiv() or itruediv() |
//= | ifloordiv() |
%= | imod() |
**= | ipow() |
<<= | ilshift() |
>>= | irshift() |
&= | iand() |
|= | ior() |
^= | ixor() |
In NumPy a function like np.sqrt(a)
can either return a scalar or an array, depending on the value passed into it. In C# we are trying to solve this by creating overloads for every scalar type in addition to the original function that takes an array and returns an array. This bloats the API however and hasn't been done for most functions and for some it isn't possible at all due to language limitations. The alternative is to cast value types to an array (NDarray can represent scalars too):
So a simple root = np.sqrt(2)
needs to be written somewhat clumsily like this in C#: var root = (double)np.sqrt((NDarray)2.0);
until overloads for every value type are added in time for your convenience. In any case, the main use case for NumPy functions is to operate on the elements of arrays, which is convenient enough:
var a = np.arange(100); // => { 0, 1, 2, ... , 98, 99 }
var roots = np.sqrt(a); // => { 0.0, 1.0, 1.414, ..., 9.899, 9.950 }
Python has native support of complex numbers, something the C# language is lacking as well. Converting complex results to and from NumPy is not implemented at all (yet). Please step forward if you want to work on this.
The function fft(...) in numpy.fft and random(...) in numpy.random had to be renamed because C# doesn't allow a member to have the same name as its enclosing class. That's why in Numpy.NET these functions have been renamed with a trailing underscore like this: np.fft.fft_(...)
Currently, Numpy.dll is targeting .NET Standard (on Windows) and packages the following binaries:
- Python 3.7: (python-3.7.3-embed-amd64.zip)
- NumPy 1.16 (numpy-1.16.3-cp37-cp37m-win_amd64.whl)
Numpy.Bare.dll is available for the Python versions 2.7, 3.5, 3.6 and 3.7 on Windows, Linux and MacOS on Nuget.
To make Numpy.dll support Linux a separate version of Python.Included packaging linux binaries of Python needs to be made and a version of Numpy.dll that packages a linux-compatible NumPy wheel. If you are interested, you may work on this issue.
Numpy.NET packages and distributes Python
, pythonnet
as well as numpy
. All these dependencies imprint their license conditions upon Numpy.NET. The C# wrapper itself is MIT License.
- Python: PSF License
- Python for .NET (pythonnet): MIT License
- NumPy: BSD License
- Numpy.NET: MIT License
- If you check
Prefer 32-bit
in your build config or build withx86
instead ofAny CPU
Numpy.NET will crash withBadFormatException
- If you have insufficient folder permissions in AppData Numpy.NET might crash. You can specify a different installpath by setting
Installer.INSTALL_PATH = "<install path>";