TType

Common interface for all typed tensors.

Typed tensors wrap a RawTensor by mapping their native memory to a n-dimensional data space allowing direct I/O access from the JVM.

Subinterfaces of TType are propagated as a generic parameter to various entities of TensorFlow to identify the type of the tensor they carry. For example, a Operand<TFloat32> is an operand which outputs a 32-bit floating point tensor. This parameter ensure type-compatibility between operands of a computation at compile-time. For example:

Ops tf = Ops.create();

 Constant<TFloat32> c1 = tf.array(2.0f, 3.0f, 2.0f);
 Constant<TFloat32> c2 = tf.array(1.0f, 2.0f, 3.0f);
 Constant<TInt32> c3 = tf.array(2, 3, 2);

 tf.math.add(c1, c2);  // OK
 tf.math.add(c1, c3);  // Compilation failure
 

Even if all typed tensors implements somehow NdArray to provide access to their data, TType deliberately does not extend directly from this interface, for the following reasons:

• Implementing NdArray at this level could only expose boxed-type accessors, which are less performant than their primitive equivalent, only exposed by subinterfaces of NdArray (e.g. FloatNdArray).
• TType would need to carry a new generic parameter for typing the NdArray, which will increase the verbosity in the signature of any method accepting or returning an instance of this interface, which is very common.