Tinkex.Types.TensorData (Tinkex v0.3.4)

View Source

Numerical tensor data for training.

Mirrors Python tinker.types.TensorData.

IMPORTANT: Only int64 and float32 dtypes are supported by the backend. This module performs aggressive casting to match Python SDK behavior.

Summary

Types

t()

Functions

from_nx(tensor)

Create TensorData from an Nx tensor.

to_nx(tensor_data)

Convert TensorData back to an Nx tensor.

tolist(tensor_data)

Return the flat data list from TensorData.

Types

t()

@type t() :: %Tinkex.Types.TensorData{
  data: [number()],
  dtype: Tinkex.Types.TensorDtype.t(),
  shape: [non_neg_integer()] | nil
}

Functions

from_nx(tensor)

@spec from_nx(Nx.Tensor.t()) :: t()

Create TensorData from an Nx tensor.

Performs aggressive type casting to match Python SDK:

  • float64 → float32 (downcast)
  • int32 → int64 (upcast)
  • unsigned → int64 (upcast)

to_nx(tensor_data)

@spec to_nx(t()) :: Nx.Tensor.t()

Convert TensorData back to an Nx tensor.

tolist(tensor_data)

@spec tolist(t()) :: [number()]

Return the flat data list from TensorData.

Provides API parity with Python's TensorData.tolist().

Examples

iex> tensor = TensorData.from_nx(Nx.tensor([1.0, 2.0, 3.0]))
iex> TensorData.tolist(tensor)
[1.0, 2.0, 3.0]