Tensor

The central type of tf-dotty is a Tensor, which takes two type parameters:

• T for the type of the values in the tensor,
• S for the shape of the tensor.

Tensor element type T

The type of elements in a Tensor is given by T. Most TensorFlow operations will take an optional argument dtype of type DataType[T] that determines the T type: for instance, using tf.float32 will result in a tensor of Float.

Tensor shape S

The S type parameter is a subtype of Shape. The Shape type is similar to an HList in the shapeless library, or to a Tuple in Dotty. Like HList and Tuple, it represents a list of types of any length. Unlike HList and Tuple, it can impose an upper bound on the type of the elements: the upper bound for Shape is a singleton integer type, Int & Singleton.

Analogously to a List, a Shape type is constructed by using Cons and Nil types, which are #: and SNil in tf-dotty. For example, the type 2 #: 3 #: SNil represents the shape of a \( 2 \times 3 \) matrix, and SNil corresponds to a scalar value.

Applications of Cons (#:) return the precise type of the Cons when dimension sizes are written as integer literals:

val shape1 /* inferred type: 32 #: 64 #: SNil */ = 32 #: 64 #: SNil

Shapes can also programmatically be generated from other shapes, while retaining type information. This is implemented as a match type to transform the type, and a function that casts to the match type.

val shape2: /* inferred type: 64 #: 32 #: SNil */ = shape1.reverse

Note that a shape of SNil corresponds to a single value.