kyonifer / koma Goto Github PK

This program:

import koma.eye

fun main(args: Array<String>)
{
    val a = eye(3)
    val b = eye(3)
    println(a == b)
}

Produces false, because it just compares whether a and b refer to the same Kotlin object.

It would be very useful to have an overload that would compare matrices in the mathematical sense. That is, two matrices are the same if they have the same dimensions and each element of one matrix is equal to the element in the other matrix in the same column and row.

Need the ability to save plots to disk. Also maybe add an option to run the plots headless, e.g. plt.show() on matplotlib.

for instance something like

A[0..5,:]

?

or do I have to do

A[0..5,0 until numCols]

?

beautiful project.

thanks.

Now that inline classes look slated for kotlin mainline (experimental in 1.3-M2), it would be very useful to implement Matrix<T> with inline classes to eliminate any overhead using koma might have had over using the underlying libs on each platform. It appears the feature currently supports a single wrapped property and interfaces (but not abstract bases), which could work for Matrix. There are some backends where a single thing isnt enough currently (e.g. native wants both a pointer to memory and a length) but we can probably refactor a bit to make every implementation wrap a single object.

Plotting multiple plots quickly can cause a threading error.

Currently on the JVM target, val a =zeros(3,3); a[1,1] will yield a boxed number because .get(...) has return type T, which forces Matrix<Double> to return java.lang.Double instead of double. This is inconvenient for performance reasons if calling get or set in a loop.

Attempting to use something like eye(2).QR() always seems to result in empty matrices being returned for QR when using EJML backend (not tested on others). It appears that a call to

decomp.decompose(this.storage.matrix)

after this line may help following docs for getQ() here.

Rather than arbitrarily assigning one of the current 'hard' standards (WALL) in the Time constructor, TimeStandard should default to OTHER. This matches better with the current practice of only checking TimeStandards if each time has been explicitly linked to a standard.

Say I have a 4x4 matrix called x:

val x = zeros(4, 4)

If I try to populate a column with a vector oriented the wrong way, I expect an IndexOutOfBoundsException but I get an IllegalArgumentException :

>>> x[0..end, 1] = mat[1, 2, 3, 4]
java.lang.IllegalArgumentException: Specified element is out of bounds: 1 0
	at org.ejml.data.DMatrixRMaj.get(DMatrixRMaj.java:262)
	at org.ejml.simple.SimpleBase.get(SimpleBase.java:654)
	at koma.matrix.ejml.EJMLMatrix.getDouble(EJMLMatrix.kt:43)
	at koma.extensions.MatrixExtensions__Extensions_matrix_DoubleKt.get(extensions_matrix_Double.kt:152)
	at koma.extensions.MatrixExtensions.get(Unknown Source)
	at koma.extensions.MatrixExtensions__Extensions_matrix_DoubleKt.setRangesDouble(extensions_matrix_Double.kt:224)
	at koma.extensions.MatrixExtensions.setRangesDouble(Unknown Source)
	at koma.extensions.MatrixExtensions__Extensions_matrix_DoubleKt.setRowRangeDouble(extensions_matrix_Double.kt:269)
	at koma.extensions.MatrixExtensions.setRowRangeDouble(Unknown Source)

Even worse, if I provide more data than is needed to fill the space in the correct direction, there is no error, allowing me to write this off-by-one code by accident:

>>> x[0 until end, 1] = mat[1,2,3,4].T
>>> x
mat[ 0.00,  1.00,  0.00,  0.00 end
     0.00,  2.00,  0.00,  0.00 end
     0.00,  3.00,  0.00,  0.00 end
     0.00,  0.00,  0.00,  0.00 ]

It seems Plotting is not available with Kotlin Native.

I'd put a note about that here and here.

Example:

60000.0 - mat[10000.0, 0.00] // res: mat[-50000.00,  -60000.00  ]

I got the following error trying to build for Kotlin Native:

Exception in thread "main" java.lang.Error: /var/folders/y7/qnnjc1q143q39hwttm_rs4gm0000gn/T/tmp7387811828594230022.c:1:10: fatal error: 'cblas.h' file not found
        at org.jetbrains.kotlin.native.interop.indexer.UtilsKt.ensureNoCompileErrors(Utils.kt:137)
        at org.jetbrains.kotlin.native.interop.indexer.IndexerKt.indexDeclarations(Indexer.kt:898)
        at org.jetbrains.kotlin.native.interop.indexer.IndexerKt.buildNativeIndexImpl(Indexer.kt:888)
        at org.jetbrains.kotlin.native.interop.indexer.NativeIndexKt.buildNativeIndex(NativeIndex.kt:56)
        at org.jetbrains.kotlin.native.interop.gen.jvm.MainKt.processCLib(main.kt:307)
        at org.jetbrains.kotlin.native.interop.gen.jvm.MainKt.interop(main.kt:38)
        at org.jetbrains.kotlin.cli.utilities.InteropCompilerKt.invokeInterop(InteropCompiler.kt:100)
        at org.jetbrains.kotlin.cli.utilities.MainKt.main(main.kt:29)

I installed openblas but it didn't help.

Now that uint will be a builtin kotlin feature, we should deprecate our uint support entirely. In addition, inline classes leave us an avenue to support complex matrices without boxing, and consequently complex algorithms (such as ffts).

Should be a supertype of 2D, without the 2D only operations (will require specific backend support, since none of our 2D backends work with N atm).

Currently koma only tests itself on the jvm. It needs to instead switch to kotlin-test and test on all 3 backends, as well as extend test coverage.

          Matrix           Required Actual 
========================== ======== ====== 
measurementMatrix             NxM     2x1  
measurementNoiseCovariance    NxN     2x2  
measurement                   Nx1     1x2  
xhat                          Mx1     2x1  
P                             MxM     2x2  
measurementMatrix must have the same number of rows as measurementNoiseCovariance has rows
measurementMatrix must have the same number of rows as measurementNoiseCovariance has columns
measurementMatrix must have the same number of rows as measurement has rows
measurement must have exactly 1 columns (has 2)

In the above example, the first two lines of the narrative should not appear because those constraints are already satisfied.

Hi kyonier, great lib!
Since dataframe support still seems missing in golem, you maybe could add a reference to
https://github.com/holgerbrandl/kplyr

In #43 equality was changed to be elementwise. This currently causes matrix validation to break in some cases. For example:

val x = zeros(3,2)
val y = zeros(3,2)
validate {
    x("x"){'N' x 1}
    y("y"){'N' x 'M'}
}

This should fail (because x must have 1 column) but doesn't.

The issue is that DimensionValidator constructs a dimensions hashmap from matrices to possible dimensions. Because y and x are equal under elementwise equality they hash to the same bin in DimensionValidator.dimensions. Thus x's dimension list is clobbered by y's and x survives the dimension check.

Its probably enough to change the validator hashes to use a referential equality wrapper.

Currently this is left to throw an unimplemented error. Ideally an e.g. EJML matrix should be multipliable by a MTJ matrix.

Hi, I saw you redefined PI an E here.
But doing so is pointless since the val you defined (could have been a const val BTW) is of type Double, which doesn't have as much precision as the digits you wrote.
Kotlin's 1.2 math package already includes these constants (which forwards to JVM constants that already have the max precision achievable for type Double).

To prove what I'm saying, here's a snippet that runs without Exception. You can try it on try.kotl.in

typealias BlaBla = Array<String>

fun main(argsBlaBla: BlaBla) {
    
    val tooLongPiForDouble = 3.1415926535897932384626433832795028841971693993751058209749445923078164062
    check(tooLongPiForDouble == 3.14159265358979323846)
    check(tooLongPiForDouble == Math.PI)
    check(Math.PI == kotlin.math.PI)
    check(kotlin.math.PI == 3.14159265358979323846)
    println(tooLongPiForDouble)
    println(kotlin.math.PI)
}

The current DefaultXXMatrix implementations currently have a lot of unimplemented methods. On some platforms (e.g. js-in-browser) finding an apache license compatible lib that has the full set of functionality can be challenging, so it would be useful to have a set of complete default implementations for full compatibility.

Split should provide isolation for:

• logging
• core api
• each backend

It can be counter-intuitive when you want to use a time with default standard which in unexpected. By default, Time without a standard should silently pass checks.

A matrix log is needed to solve A in e^A=B (i.e. not an element wise log). It doesn't look like this method is available under any of the current dependencies, but is available under jeigen.

https://github.com/hughperkins/jeigen#matrix-exponential-matrix-logarithm

Create a matrix division operator to solve Ax=b which avoids computing the full inverse of A.

Would it be possible to provide a public (e.g. Jcenter/bintray) artifact id for golem?

Currently the Matrix interface has times() operator for multiplying by both other matrices and numbers.
Hence, this code is correct:
var mulMatByTwo = ones(1) * 2.0

But as times() operator is not extended in the Int and Double classes, such code is incorrect:
var mulTwoByMat = 2.0 * ones(1)
which means multiplication is not commutative.

For the sake of consistency, we can create an extension in the Number class:
operator fun <M : Number> Number.times(other: Matrix<M>): Matrix<Double> { return other * this }

NDArray's map is defined as fun map(f: (T) -> T): NDArray<T>. It would be more useful if map were a generic, in the sense of fun map<R>(f: (T) -> R): NDArray<R>.

Here's some sample code which I'd like to compile, but can't:

var counter = 0
val ints: NDArray<Int> = DefaultNDArray(3, 4, 5) { ++counter }
val doubles: NDArray<Double> = ints.map { it.toDouble() }

Currently I'm using the following hacky workaround, which is dangerous and insane but seems to work.

fun <TIn, TOut> NDArray<TIn>.crazyMapWorkaround(fn: (TIn) -> TOut) = 
    toIterable().iterator().let { itr ->
        DefaultNDArray(*(shape().toIntArray())) { fn(itr.next()) }
    }

On adding the repository in the Gradle I am getting the following error.

Java: file not found Exception
Resource nexus-maven-repository-index-properties does not exist

Command ./gradlew build produces this error on mac

Task :koma-core-cblas:compileKonanCblasMacbook FAILED
Exception in thread "main" java.lang.Error: /var/folders/wz/xhlg5g6n1c785tsbryqp0yyh0000gn/T/tmp7098207755916076298.c:1:10: fatal error: 'cblas.h' file not found
        at org.jetbrains.kotlin.native.interop.indexer.UtilsKt.ensureNoCompileErrors(Utils.kt:137)
        at org.jetbrains.kotlin.native.interop.indexer.IndexerKt.indexDeclarations(Indexer.kt:898)
        at org.jetbrains.kotlin.native.interop.indexer.IndexerKt.buildNativeIndexImpl(Indexer.kt:888)
        at org.jetbrains.kotlin.native.interop.indexer.NativeIndexKt.buildNativeIndex(NativeIndex.kt:56)
        at org.jetbrains.kotlin.native.interop.gen.jvm.MainKt.processCLib(main.kt:307)
        at org.jetbrains.kotlin.native.interop.gen.jvm.MainKt.interop(main.kt:38)
        at org.jetbrains.kotlin.cli.utilities.InteropCompilerKt.invokeInterop(InteropCompiler.kt:100)
        at org.jetbrains.kotlin.cli.utilities.MainKt.main(main.kt:29)

I try to modify
buildscripts/cblas.def into

headers=cblas.h
compilerOpts = -I/usr/include/x86_64-linux-gnu -I/usr/local/opt/openblas/include
linkerOpts = -L/usr/lib -L/usr/lib/x86_64-linux-gnu -L/usr/local/opt/openblas/lib -lblas

and
buildscripts/lapacke.def into

headers=lapacke.h
compilerOpts = -I/usr/local/opt/lapack/include
linkerOpts = -L/usr/lib -L/usr/lib/x86_64-linux-gnu -L/usr/local/opt/lapack/lib -llapacke

but no result, any idea?

For the js side, we may end up needing two backends: one for node (which calls out to optimized blas, see e.g. Vectorious) and one for everything else (browser, etc). One route that might enable us to get the "everything else" bin implementing all of our ops could be the automatic emscripten blas generators that are out there.

This page on the web site:
https://koma.kyonifer.com/General_Usage_Guide/Advanced/Build_From_Source/index.html

Says the following command generates the KLib:

./gradlew compileKonanKoma -Ptarget=native

But the right command is:

./gradlew compileKonanKomaKLib -Ptarget=native

Just need to replace compileKonanKoma with compileKonanKomaKLib.

After switching to a unified build system using multiplatform, the build docs also need to be updated. The new build steps are buildNative, buildJvm, and buildJs, with build running all of them.

You can set columns (and I'd guess rows, I haven't checked) with an arbitrary Matrix that has fewer than the expected number of elements in the dimension to be set, and the fill data is taken from the rows of the matrix passed in up to the number of cols available:

val m = zeros(3, 2)
val fillVal = ones(2, 1000).fill{r, c -> c + 1.0} // 2 x 2 works also
m.setCol(0, fillVal) 
// The transpose still fails on size check
// m.setCol(0, fillVal.T) 

// Now m.getCol(0) returns
mat[ 1.00 end
     2.00 end
     0.00 ]