Matft is Numpy-like library in Swift. Function name and usage is similar to Numpy.

• The MfArray such like a numpy.ndarray

  let a = MfArray([[[ -8,  -7,  -6,  -5],
                    [ -4,  -3,  -2,  -1]],
          
                   [[ 0,  1,  2,  3],
                    [ 4,  5,  6,  7]]])
  let aa = Matft.arange(start: -8, to: 8, by: 1, shape: [2,2,4])
  print(a)
  print(aa)
  /*
  mfarray = 
  [[[	-8,		-7,		-6,		-5],
  [	-4,		-3,		-2,		-1]],
  
  [[	0,		1,		2,		3],
  [	4,		5,		6,		7]]], type=Int, shape=[2, 2, 4]
  mfarray = 
  [[[	-8,		-7,		-6,		-5],
  [	-4,		-3,		-2,		-1]],
  
  [[	0,		1,		2,		3],
  [	4,		5,		6,		7]]], type=Int, shape=[2, 2, 4]
  */

• You can pass MfType as MfArray's argument mftype: .Hoge . It is similar to dtype.

  ※Note that stored data type will be Float or Double only even if you set MfType.Int. So, if you input big number to MfArray, it may be cause to overflow or strange results in any calculation (+, -, *, /,... etc.). But I believe this is not problem in practical use.

• MfType's list is below

    public enum MfType: Int{
      case None // Unsupportted
      case Bool
      case UInt8
      case UInt16
      case UInt32
      case UInt64
      case UInt
      case Int8
      case Int16
      case Int32
      case Int64
      case Int
      case Float
      case Double
      case Object // Unsupported
  }

• Also, you can convert MfType easily using astype

  let a = MfArray([[[ -8,  -7,  -6,  -5],
                    [ -4,  -3,  -2,  -1]],
          
                   [[ 0,  1,  2,  3],
                    [ 4,  5,  6,  7]]])
  print(a)//See above. if mftype is not passed, MfArray infer MfType. In this example, it's MfType.Int
  
  let a = MfArray([[[ -8,  -7,  -6,  -5],
                    [ -4,  -3,  -2,  -1]],
              
                   [[ 0,  1,  2,  3],
                    [ 4,  5,  6,  7]]], mftype: .Float)
  print(a)
  /*
  mfarray = 
  [[[	-8.0,		-7.0,		-6.0,		-5.0],
  [	-4.0,		-3.0,		-2.0,		-1.0]],
  
  [[	0.0,		1.0,		2.0,		3.0],
  [	4.0,		5.0,		6.0,		7.0]]], type=Float, shape=[2, 2, 4]
  */
  let aa = MfArray([[[ -8,  -7,  -6,  -5],
                    [ -4,  -3,  -2,  -1]],
              
                   [[ 0,  1,  2,  3],
                    [ 4,  5,  6,  7]]], mftype: .UInt)
  print(aa)
  /*
  mfarray = 
  [[[	4294967288,		4294967289,		4294967290,		4294967291],
  [	4294967292,		4294967293,		4294967294,		4294967295]],
  
  [[	0,		1,		2,		3],
  [	4,		5,		6,		7]]], type=UInt, shape=[2, 2, 4]
  */
  //Above output is same as numpy!
  /*
  >>> np.arange(-8, 8, dtype=np.uint32).reshape(2,2,4)
  array([[[4294967288, 4294967289, 4294967290, 4294967291],
          [4294967292, 4294967293, 4294967294, 4294967295]],
  
         [[         0,          1,          2,          3],
          [         4,          5,          6,          7]]], dtype=uint32)
  */
  
  print(aa.astype(.Float))
  /*
  mfarray = 
  [[[	-8.0,		-7.0,		-6.0,		-5.0],
  [	-4.0,		-3.0,		-2.0,		-1.0]],
  
  [[	0.0,		1.0,		2.0,		3.0],
  [	4.0,		5.0,		6.0,		7.0]]], type=Float, shape=[2, 2, 4]
  */

• You can access specific data using subscript.

You can set MfSlice (see below's list) to subscript.

• MfSlice(start: Int? = nil, to: Int? = nil, by: Int = 1)

• Matft.newaxis

• ~< //this is prefix, postfix and infix operator. same as python's slice, ":"

• Normal indexing

  let a = Matft.arange(start: 0, to: 27, by: 1, shape: [3,3,3])
  print(a)
  /*
  mfarray = 
  [[[	0,		1,		2],
  [	3,		4,		5],
  [	6,		7,		8]],
  
  [[	9,		10,		11],
  [	12,		13,		14],
  [	15,		16,		17]],
  
  [[	18,		19,		20],
  [	21,		22,		23],
  [	24,		25,		26]]], type=Int, shape=[3, 3, 3]
  */
  print(a[2,1,0])
  // 21

  Slicing

• If you replace : with ~<, you can get sliced mfarray. Note that use a[0~<] instead of a[:] to get all elements along axis.

  print(a[~<1])  //same as a[:1] for numpy
  /*
  mfarray = 
  [[[	9,		10,		11],
  [	12,		13,		14],
  [	15,		16,		17]]], type=Int, shape=[1, 3, 3]
  */
  print(a[1~<3]) //same as a[1:3] for numpy
  /*
  mfarray = 
  [[[	9,		10,		11],
  [	12,		13,		14],
  [	15,		16,		17]],
  
  [[	18,		19,		20],
  [	21,		22,		23],
  [	24,		25,		26]]], type=Int, shape=[2, 3, 3]
  */
  print(a[~<~<2]) //same as a[::2] for numpy
  //print(a[~<<2]) //alias
  /*
  mfarray = 
  [[[	0,		1,		2],
  [	3,		4,		5],
  [	6,		7,		8]],
  
  [[	18,		19,		20],
  [	21,		22,		23],
  [	24,		25,		26]]], type=Int, shape=[2, 3, 3]
  */

• Negative indexing is also available That's implementation was hardest for me...

  print(a[~<-1])
  /*
  mfarray = 
  [[[	0,		1,		2],
  [	3,		4,		5],
  [	6,		7,		8]],
  
  [[	9,		10,		11],
  [	12,		13,		14],
  [	15,		16,		17]]], type=Int, shape=[2, 3, 3]
  */
  print(a[-1~<-3])
  /*
  mfarray = 
  	[], type=Int, shape=[0, 3, 3]
  */
  print(a[~<~<-1])
  //print(a[~<<-1]) //alias
  /*
  mfarray = 
  [[[	18,		19,		20],
  [	21,		22,		23],
  [	24,		25,		26]],
  
  [[	9,		10,		11],
  [	12,		13,		14],
  [	15,		16,		17]],
  
  [[	0,		1,		2],
  [	3,		4,		5],
  [	6,		7,		8]]], type=Int, shape=[3, 3, 3]*/

• Note that returned subscripted mfarray will have base property (is similar to view in Numpy). See numpy doc in detail.

  let a = Matft.arange(start: 0, to: 4*4*2, by: 1, shape: [4,4,2])
              
  let b = a[0~<, 1]
  b[~<<-1] = MfArray([9999]) // cannot pass Int directly such like 9999
  
  print(a)
  /*
  mfarray = 
  [[[	0,		1],
  [	9999,		9999],
  [	4,		5],
  [	6,		7]],
  
  [[	8,		9],
  [	9999,		9999],
  [	12,		13],
  [	14,		15]],
  
  [[	16,		17],
  [	9999,		9999],
  [	20,		21],
  [	22,		23]],
  
  [[	24,		25],
  [	9999,		9999],
  [	28,		29],
  [	30,		31]]], type=Int, shape=[4, 4, 2]
  */

Below is Matft's function list. As I mentioned above, almost functions are similar to Numpy. Also, these function use Accelerate framework inside, the perfomance may keep high.

* means method function exists too. Shortly, you can use a.shallowcopy() when a is MfArray.

• Creation

• Conversion

• File save function has not developed yet.

• Operation

  Line 2 is infix (prefix) operator.

• Math function

Other function is also available. See here.

• Statistics function

• Linear algebra

I use Accelerate, so all of MfArray operation may keep high performance.

func testPefAdd1() {
        do{
            let a = Matft.arange(start: 0, to: 10*10*10*10*10*10, by: 1, shape: [10,10,10,10,10,10])
            let b = Matft.arange(start: 0, to: -10*10*10*10*10*10, by: -1, shape: [10,10,10,10,10,10])
            
            self.measure {
                let _ = a+b
            }
            /*
             '-[MatftTests.ArithmeticPefTests testPefAdd1]' measured [Time, seconds] average: 0.001, relative standard deviation: 23.418%, values: [0.001707, 0.001141, 0.000999, 0.000969, 0.001029, 0.000979, 0.001031, 0.000986, 0.000963, 0.001631]
            1.14ms
             */
        }
    }
    
    func testPefAdd2(){
        do{
            let a = Matft.arange(start: 0, to: 10*10*10*10*10*10, by: 1, shape: [10,10,10,10,10,10])
            let b = a.transpose(axes: [0,3,4,2,1,5])
            let c = a.T
            
            self.measure {
                let _ = b+c
            }
            /*
             '-[MatftTests.ArithmeticPefTests testPefAdd2]' measured [Time, seconds] average: 0.004, relative standard deviation: 5.842%, values: [0.004680, 0.003993, 0.004159, 0.004564, 0.003955, 0.004200, 0.003998, 0.004317, 0.003919, 0.004248]
            4.20ms
             */
        }
    }

    func testPefAdd3(){
        do{
            let a = Matft.arange(start: 0, to: 10*10*10*10*10*10, by: 1, shape: [10,10,10,10,10,10])
            let b = a.transpose(axes: [1,2,3,4,5,0])
            let c = a.T
            
            self.measure {
                let _ = b+c
            }
            /*
             '-[MatftTests.ArithmeticPefTests testPefAdd3]' measured [Time, seconds] average: 0.004, relative standard deviation: 16.815%, values: [0.004906, 0.003785, 0.003702, 0.005981, 0.004261, 0.003665, 0.004083, 0.003654, 0.003836, 0.003874]
            4.17ms
             */
        }

Matft achieved almost same performance as Numpy!!!

※Swift's performance test was conducted in release mode

My codes have several overhead and redundant part so this performance could be better than now.

import numpy as np
#import timeit

a = np.arange(10**6).reshape((10,10,10,10,10,10))
b = np.arange(0, -10**6, -1).reshape((10,10,10,10,10,10))

#timeit.timeit("b+c", repeat=10, globals=globals())
%timeit -n 10 a+b
"""
962 µs ± 273 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
"""

a = np.arange(10**6).reshape((10,10,10,10,10,10))
b = a.transpose((0,3,4,2,1,5))
c = a.T
#timeit.timeit("b+c", repeat=10, globals=globals())
%timeit -n 10 b+c
"""
5.68 ms ± 1.45 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
"""

a = np.arange(10**6).reshape((10,10,10,10,10,10))
b = a.transpose((1,2,3,4,5,0))
c = a.T
#timeit.timeit("b+c", repeat=10, globals=globals())
%timeit -n 10 b+c
"""
3.92 ms ± 897 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
"""

• Import
  • Project > Build Setting > +
  • Select Rules
• Update
  • File >Swift Packages >Update to Latest Package versions

• Set Cartfile

  echo 'github "jjjkkkjjj/Matft"' > Cartfile
   carthage update ###or append '--platform ios'
  

• Import Matft.framework made by above process to your project

• Create Podfile (Skip if you have already done)

  pod init

• Write pod 'Matft' in Podfile such like below

  target 'your project' do
    pod 'Matft'
  end

• Install Matft

  pod install