Tensor

A named, shaped block of numbers: one input to or output from a model. Elements are flat and row-major; the shape says how to read them. A 1x3x224x224 image is 150,528 numbers with shape [1,3,224,224], and nesting it would cost a Vector per row for nothing. Floats and integers are held SEPARATELY rather than one being converted to the other. Both are needed in practice -- pixels and embeddings are FP32, token ids for a language model are INT64 -- and a conversion at construction would impose the narrower type's limits on a backend that does not have them. An HTTP server receives an integer tensor as JSON integers with no range limit at all; only the native crossing needs floats, so that is where the conversion and its check live.

Example

pixels := Float->New[1 * 3 * 224 * 224];
input := Tensor->New("data", [1, 3, 224, 224], pixels);

ids := [101, 2003, 1037];
tokens := Tensor->New("input_ids", [1, 3], ids);

Operations

• New
• AsFloats
• GetCount
• GetDatatype
• GetFloats
• GetInts
• GetName
• GetShape
• GetShapeProduct
• IsConsistent
• IsFloat

method : public : AsFloats() ~ Float[]

method : public : GetCount() ~ Int

method : public : GetDatatype() ~ String

method : public : GetFloats() ~ Float[]

method : public : GetInts() ~ Int[]

method : public : GetName() ~ String

method : public : GetShape() ~ Int[]

method : public : GetShapeProduct() ~ Int

method : public : IsConsistent() ~ Bool

method : public : IsFloat() ~ Bool

New(name:String, shape:Int[], data:Float[])

New(name:String, shape:Int[], data:Int[])

New(name:String, datatype:String, shape:Int[], data:Float[])