NdArray is a library exposing utilities for manipulating data in a n-dimensional space in Java. Unlike other Java artifacts distributed by TensorFlow, this library does not depend on the TensorFlow runtime, therefore is very lightweight and can be used by any kind of Java project.
To import the NdArray library in your project, simply add the following dependency:
<dependency> <groupId>org.tensorflow</groupId> <artifactId>ndarray</artifactId> <version>1.0.0</version> </dependency>Instances of DataBuffer map contiguous segments of memory with 64-bits indexing and supports generic parametrization while still allowing direct access to primitive types. Such segments could be standard Java arrays, JDK NIO buffers or native memory. In addition, it can serialize and deserialize data of any type (and not only primitive types, as with java.util.nio).
// Allocate a buffer of 4K int valuesIntDataBufferbufferA = DataBuffers.ofInts(4096L); assertEquals(4096L, bufferA.size()); // Write an int array at the beginning of the bufferbufferA.write(newint[]{1, 2, 3 }); assertEquals(3, bufferA.getInt(2)); // Slice buffer after its first valueIntDataBufferbufferB = bufferA.offset(1); assertEquals(4095L, bufferB.size()); assertEquals(2, bufferB.getInt(0)); // Resize a buffer to 10 elementsIntDataBufferbufferC = bufferB.narrow(10); assertEquals(10L, bufferB.size()); assertEquals(2, bufferB.getInt(0));Instances of NdArray are used to view memory segments stored in a DataBuffer as a multidimensional arrays and to provide an API for traversing, reading, writing and slicing their data. The goal of these tools is to replace the usage of standard multidimensional Java arrays (e.g. new int[][][]) since those results in slow performances, from the non-contiguous storage of their data and the multiple dereferences required to access their values.
// Allocating a 3D matrix of 2x3x2IntNdArraymatrix3d = NdArrays.ofInts(Shape.of(2, 3, 2)); assertEquals(3, matrix3d.rank()); // Initializing 3D matrix data with vectors from the first dimension (index 0)matrix3d.elements(0).forEach(matrix ->{assertEquals(2, matrix.rank()); assertEquals(Shape.of(3, 2), matrix.shape()); matrix .set(NdArrays.vectorOf(1, 2), 0) .set(NdArrays.vectorOf(3, 4), 1) .set(NdArrays.vectorOf(5, 6), 2)}); // Visit all scalars of 3D matrix, printing their coordinates and valuematrix3d.scalars().forEachIdx((coords, scalar) -> System.out.println("Scalar at " + Arrays.toString(coords) + " has value " + scalar.getInt()) ); // Retrieving the second vector of the first matrixIntNdArrayvector = matrix3d.get(0, 1); assertEquals(1, vector.rank()); // Rewriting the values of the vector using a primitive arrayvector.copyFrom(DataBuffers.of(newint[]{7, 8 })); assertEquals(7, matrix3d.getInt(0, 1, 0)); assertEquals(8, matrix3d.getInt(0, 1, 1)); // Slicing the 3D matrix so we only keep the second element of the second dimensionIntNdArrayslice = matrix3d.slice(all(), at(1)); assertEquals(2, slice.rank()); assertEquals(Shape.of(2, 2), slice.shape()); assertEquals(7, slice.getInt(0, 0)); // (0, 1, 0) in the original matrixassertEquals(3, slice.getInt(1, 0)); // (1, 1, 0) in the original matrixThe NdArray library is independent of the TensorFlow runtime library, making it a good choice for manipulating multi-dimensional data structures from anywhere. But as an example, here is how it is actually being used by the TensorFlow Java API:
// Allocate a tensor of 32-bits integer of the shape (2, 3, 2)TInt32tensor = TInt32.ofShape(2, 3, 2); // Access tensor memory directlyassertEquals(3, tensor.rank()); assertEquals(12, tensor.size()); try (EagerSessionsession = EagerSession.create()){Opstf = Ops.create(session); // Initialize tensor memory with zeros and take a snapshottensor.scalars().forEach(scalar -> scalar.setInt(0)); Constant<T> x = tf.constant(tensor); // Initialize the same tensor memory with ones and take a snapshottensor.scalars().forEach(scalar -> scalar.setInt(1)); Constant<T> y = tf.constant(tensor); // Subtract y from x and validate the resultSub<T> sub = tf.math.sub(x, y); sub.asTensor().scalars().forEach(scalar -> assertEquals(-1, scalar.getInt()) )}