Type Conversions#

Daft to Python#

This table shows the mapping from Daft DataTypes to Python types, as done in places such as Series.to_pylist, Expression.cast to Python type, and arguments passed into functions decorated with @daft.func.

For Map[K, V] conversions, Daft defaults to list[tuple[K, V]] to preserve duplicate keys and ordering. If you pass maps_as_pydicts="lossy" or maps_as_pydicts="strict", Daft converts maps to Python dicts: - "lossy" keeps the last value for duplicate keys and emits a warning when duplicates are encountered. - "strict" raises an exception when duplicate keys are encountered.

Python to Daft#

From Python Type#

This table shows the mapping from Python types to Daft types, such as when inferring the return type from the type hints of a function decorated with @daft.func.

To check the inferred DataType for a Python type, use DataType.infer_from_type.

jaxtyping#

The jaxtyping library provides data type and shape type annotations for array/tensor types from various libraries, including NumPy, PyTorch, TensorFlow, and JAX. Daft is able to natively infer the inner dtype and shape from jaxtyping types.

Examples:

• jaxtyping.Float64[jaxtyping.Array, "1 2 3 4"] -> FixedShapeTensor[Float64, [1, 2, 3, 4]]
• jaxtyping.Int8[torch.Tensor, "dim1 dim2"] -> Tensor[Int8]

Dtype Inference#

The following table show the mapping from jaxtyping types to Daft DataType. The Daft DataType corresponds to the inner type of the result Tensor or FixedShapeTensor.

Shape Inference#

The second generic parameter of a jaxtyping type is a string of space-separated symbols representing the shape of the array. Daft will attempt to infer the tensor shape from the string.

• If all dimensions are fixed-size, Daft will infer a FixedShapeTensor with those dimensions.
  • E.g. "1 2 3", "rows=4 cols=3", or "" (scalar shape)
• Otherwise, Daft will infer a Tensor.
  • E.g. "dim1 dim2", "512 512 _", or "... 1 2 3"

From Python Object#

In addition to the above table, this table shows the additional behavior when Daft converts Python objects to Daft types without an explicitly specified type, such as in daft.from_pydict and Series.from_pylist. In these cases, Daft is able to derive information from the Python object that is not present in the object's type, allowing for better mapping to Daft types.

To check the inferred DataType for a Python object, use DataType.infer_from_object.