Module org.tensorflow.ndarray
Package org.tensorflow.ndarray.impl.sparse

Class AbstractSparseNdArray<T,​U extends NdArray<T>>

    • Constructor Detail

      • AbstractSparseNdArray

        protected AbstractSparseNdArray​(LongNdArray indices,
                                U values,
                                T defaultValue,
                                DimensionalSpace dimensions)
        Creates an abstract SparseNdArray
        Parameters:
        indices - A 2-D LongNdArray of shape [N, ndims], that specifies the indices of the elements in the sparse array that contain non-default values (elements are zero-indexed). For example, indices=[[1,3], [2,4]] specifies that the elements with indexes of [1,3] and [2,4] have non-default values.
        values - A 1-D NdArray of any type and shape [N], which supplies the values for each element in indices. For example, given indices=[[1,3], [2,4]], the parameter values=[18, 3.6] specifies that element [1,3] of the sparse NdArray has a value of 18, and element [2,4] of the NdArray has a value of 3.6.
        defaultValue - Scalar value to set for indices not specified in indices
        dimensions - the dimensional space for the dense object represented by this sparse array.

      • AbstractSparseNdArray

        protected AbstractSparseNdArray​(T defaultValue,
                                DimensionalSpace dimensions)
        Creates an abstract SparseNdArray
        Parameters:
        defaultValue - Scalar value to set for indices not specified in getIndices()
        dimensions - the dimensional space for the dense object represented by this sparse array,

    • Method Detail

      • elements

        public NdArraySequence<U> elements​(int dimensionIdx)
        Returns a sequence of all elements at a given dimension.

        Logically, the N-dimensional array can be flatten in a single vector, where the scalars of the (n - 1)th element precedes those of the (n)th element, for a total of Shaped.size() values.

        For example, given a n x m matrix on the [x, y] axes, elements are iterated in the following order:

        x0y0, x0y1, ..., x0ym-1, x1y0, x1y1, ..., xn-1ym-1

        The returned sequence can then be iterated to visit each elements, either by calling Iterable.forEach(Consumer) or NdArraySequence.forEachIndexed(BiConsumer). 

        
    // Iterate matrix for initializing each of its vectors
    matrixOfFloats.elements(0).forEach(v -> {
      v.set(vector(1.0f, 2.0f, 3.0f));
    });

    // Iterate a vector for reading each of its scalar
    vectorOfFloats.scalars().forEachIdx((coords, s) -> {
      System.out.println("Value " + s.getFloat() + " found at " + coords);
    });
        Specified by:
        elements in interface NdArray<T>
        Parameters:
        dimensionIdx - index of the dimension
        Returns:
        an NdArray sequence

      • toCoordinates

        protected long[] toCoordinates​(DimensionalSpace dimensions,
                               long position)
        Computes the coordinates based on a relative position to the beginning of the dimension space.
        Parameters:
        dimensions - the dimension space
        position - relative position to the beginning of the dimension space.
        Returns:
        the coordinates

      • getIndicesCoordinates

        protected long[] getIndicesCoordinates​(LongNdArray l)
        Converts the given set of indices coordinates to a long array of coordinates.

        The shape of the NdArray is [ndims]

        Parameters:
        l - the LongNdArray containing the coordinates
        Returns:
        the long array containing the coordinates.

      • toDense

        public abstract U toDense()
        Converts this sparse array to a dense array.
        Returns:
        the dense array.

      • slice

        public NdArray<T> slice​(Index... indices)
        Creates a multi-dimensional view (or slice) of this array by mapping one or more dimensions to the given index selectors.

        Slices allow to traverse an N-dimensional array in any of its axis and/or to filter only elements of interest. For example, for a given matrix on the [x, y] axes, it is possible to iterate elements at y=0 for all x.

        Any changes applied to the returned slice affect the data of this array as well, as there is no copy involved.

        Example of usage: 

        
    FloatNdArray matrix3d = NdArrays.ofFloats(shape(3, 2, 4));  // with [x, y, z] axes

    // Iterates elements on the x axis by preserving only the 3rd value on the z axis,
    // (i.e. [x, y, 2])
    matrix3d.slice(all(), all(), at(2)).elements(0).forEach(m -> {
      assertEquals(shape(2), m); // y=2, z=0 (scalar)
    });

    // Creates a slice that contains only the last element of the y axis and elements with an
    // odd `z` coordinate.
    FloatNdArray slice = matrix3d.slice(all(), at(1), odd());
    assertEquals(shape(3, 2), slice.shape());  // x=3, y=0 (scalar), z=2 (odd coordinates)

    // Iterates backward the elements on the x axis
    matrix3d.slice(flip()).elements(0).forEach(m -> {
      assertEquals(shape(2, 4), m);  // y=2, z=4
    });
        Specified by:
        slice in interface NdArray<T>
        Parameters:
        indices - index selectors per dimensions, starting from dimension 0 of this array.
        Returns:
        the element resulting of the index selection

      • get

        public NdArray<T> get​(long... coordinates)
        Returns the N-dimensional element of this array at the given coordinates.

        Elements of any of the dimensions of this array can be retrieved. For example, if the number of coordinates is equal to the number of dimensions of this array, then a rank-0 (scalar) array is returned, which value can then be obtained by calling `array.getObject()`.

        Any changes applied to the returned elements affect the data of this array as well, as there is no copy involved.

        Note that invoking this method is an equivalent and more efficient way to slice this array on single scalar, i.e. array.get(x, y, z) is equal to array.slice(at(x), at(y), at(z))

        Specified by:
        get in interface NdArray<T>
        Parameters:
        coordinates - coordinates of the element to access, none will return this array
        Returns:
        the element at this index

      • getObject

        public T getObject​(long... coordinates)
        Returns the value of the scalar found at the given coordinates.

        To access the scalar element, the number of coordinates provided must be equal to the number of dimensions of this array (i.e. its rank). For example: 

        
  FloatNdArray matrix = NdArrays.ofFloats(shape(2, 2));  // matrix rank = 2
  matrix.getObject(0, 1);  // succeeds, returns 0.0f
  matrix.getObject(0);  // throws IllegalRankException

  FloatNdArray scalar = matrix.get(0, 1);  // scalar rank = 0
  scalar.getObject();  // succeeds, returns 0.0f
        Note: if this array stores values of a primitive type, prefer the usage of the specialized method in the subclass for that type. For example, floatArray.getFloat(0); .
        Specified by:
        getObject in interface NdArray<T>
        Parameters:
        coordinates - coordinates of the scalar to resolve
        Returns:
        value of that scalar

      • setObject

        public NdArray<T> setObject​(T value,
                            long... coords)
        Assigns the value of the scalar found at the given coordinates.

        To access the scalar element, the number of coordinates provided must be equal to the number of dimensions of this array (i.e. its rank). For example: 

        
  FloatNdArray matrix = NdArrays.ofFloats(shape(2, 2));  // matrix rank = 2
  matrix.setObject(10.0f, 0, 1);  // succeeds
  matrix.setObject(10.0f, 0);  // throws IllegalRankException

  FloatNdArray scalar = matrix.get(0, 1);  // scalar rank = 0
  scalar.setObject(10.0f);  // succeeds
        Note: if this array stores values of a primitive type, prefer the usage of the specialized method in the subclass for that type. For example, floatArray.setFloat(10.0f, 0);
        Specified by:
        setObject in interface NdArray<T>
        Parameters:
        value - the value to assign
        coords - coordinates of the scalar to assign
        Returns:
        this array

      • set

        public NdArray<T> set​(NdArray<T> src,
                      long... coordinates)
        Assigns the value of the N-dimensional element found at the given coordinates.

        The number of coordinates provided can be anywhere between 0 and rank - 1. For example: 

        
  FloatNdArray matrix = NdArrays.ofFloats(shape(2, 2));  // matrix rank = 2
  matrix.set(vector(10.0f, 20.0f), 0);  // success
  matrix.set(scalar(10.0f), 1, 0); // success
        Specified by:
        set in interface NdArray<T>
        Parameters:
        src - an array of the values to assign
        coordinates - coordinates of the element to assign
        Returns:
        this array

      • createValues

        public abstract U createValues​(Shape shape)
        Creates a dense array of the type that this sparse array represents.
        Parameters:
        shape - the shape of the dense array.
        Returns:
        the dense of the type that this sparse array represents.

      • copyTo

        public NdArray<T> copyTo​(NdArray<T> dst)
        Copy the content of this array to the destination array.

        The Shaped.shape() of the destination array must be equal to the shape of this array, or an exception is thrown. After the copy, the content of both arrays can be altered independently, without affecting each other.

        Specified by:
        copyTo in interface NdArray<T>
        Parameters:
        dst - array to receive a copy of the content of this array
        Returns:
        this array

      • positionOf

        protected long positionOf​(long[] coords,
                          boolean isValue)
        Computes the position within the dense array given by the coordinates
        Parameters:
        coords - the coordinates within the dense array
        isValue - indicator whether the coordinates represents a value or higher level dimension.
        Returns:
        the position within the array

      • slowCopyTo

        protected void slowCopyTo​(NdArray<T> array)
        Overrides:
        slowCopyTo in class org.tensorflow.ndarray.impl.AbstractNdArray<T,​U extends NdArray<T>>

      • setIndices

        public void setIndices​(LongNdArray indices)
        Sets the Indices
        Parameters:
        indices - the Indices

      • setValues

        public void setValues​(U values)
        Sets the values
        Parameters:
        values - the values

      • locateIndex

        protected long locateIndex​(long[] coordinates)
        Gets the values index by coordinates
        Parameters:
        coordinates - the coordinates to locate
        Returns:
        index of the coordinates, if the coordinates are contained in the indices array; otherwise, (-(insertion point) - 1). The insertion point is defined as the point at which the coordinates would be inserted into the indices array: the index of the first element greater than the key, or indices.shape().get(0); if all elements in the array are less than the specified key. Note that this guarantees that the return value will be >= 0, only if the coordinates are found.

      • hashCode

        public int hashCode()
        Overrides:
        hashCode in class org.tensorflow.ndarray.impl.AbstractNdArray<T,​U extends NdArray<T>>

      • equals

        public boolean equals​(Object obj)
        Checks equality between n-dimensional arrays.

        An array is equal to another object if this object is another NdArray of the same shape, type and the elements are equal and in the same order. For example: 

        
 IntNdArray array = NdArrays.ofInts(Shape.of(2, 2))
    .set(NdArrays.vectorOf(1, 2), 0)
    .set(NdArrays.vectorOf(3, 4), 1);

 assertEquals(array, StdArrays.ndCopyOf(new int[][] {{1, 2}, {3, 4}}));  // true
 assertEquals(array, StdArrays.ndCopyOf(new Integer[][] {{1, 2}, {3, 4}}));  // true, as Integers are equal to ints
 assertNotEquals(array, NdArrays.vectorOf(1, 2, 3, 4));  // false, different shapes
 assertNotEquals(array, StdArrays.ndCopyOf(new int[][] {{3, 4}, {1, 2}}));  // false, different order
 assertNotEquals(array, StdArrays.ndCopyOf(new long[][] {{1L, 2L}, {3L, 4L}}));  // false, different types

        Note that the computation required to verify equality between two arrays can be expensive in some cases and therefore, it is recommended to not use this method in a critical path where performances matter.

        Specified by:
        equals in interface NdArray<T>
        Overrides:
        equals in class org.tensorflow.ndarray.impl.AbstractNdArray<T,​U extends NdArray<T>>
        Parameters:
        obj - object to compare this array with
        Returns:
        true if this array is equal to the provided object

      • toString

        public String toString()
        A String showing the type, default value, number of elements and the dense shape of this sparse ndarray.
        Overrides:
        toString in class Object
        Returns:
        A string containing the type, default value, number of elements and shape.

      • sortIndicesAndValues

        public AbstractSparseNdArray<T,​U> sortIndicesAndValues()
        Sorts the indices and values in ascending row-major coordinates.
        Returns:
        this instance

      • getDefaultValue

        public T getDefaultValue()
        Scalar value to set for indices not specified in indices, defaults to zero, false, or the empty String depending on the data type.

      • setDefaultValue

        public void setDefaultValue​(T defaultValue)
        Sets the defaultValue
        Parameters:
        defaultValue - the default value

      • createDefaultArray

        public abstract U createDefaultArray()
        Creates the NdArray with the default value as a scalar
        Returns:
        the default NdArray of the default value as a scalar

      • getDefaultArray

        public U getDefaultArray()
        Scalar NdArray to use for indices not specified in getIndices() This will default to zero, false, or the empty string depending on the data type of the values, otherwise it will contain the defaultValue.