tensorbox

tensorbox

tensorbox allows you to interact with dataclasses of tensors as if they were tensors. Simply use @tensorbox instead of @dataclass.

from jaxtyping import Float
from tensorbox import tensorbox
from torch import Tensor

# Define a @tensorbox class. The jaxtyping annotations describe each attribute's scalar (unbatched) shape.
@tensorbox
class Gaussians:
    mean: Float[Tensor, "dim"]
    covariance: Float[Tensor, "dim dim"]
    color: Float[Tensor, "3"]

# Define Gaussians with batch size (10, 10) and dim=3.
gaussians = Gaussians(
    torch.zeros((10, 10, 3), dtype=torch.float32),
    torch.zeros((10, 10, 3, 3), dtype=torch.float32),
    torch.zeros((10, 10, 3), dtype=torch.float32),
)

# Define a function that uses Gaussians as input. When a @tensorbox class is subscripted, each attribute's shape becomes the concatenation of the subscript (batch shape) and the attribute's original (scalar) shape. This means fn expects the following shapes:
# - mean: "batch_a batch_b dim"
# - covariances: "batch_a batch_b dim dim"
# - color: "batch_a batch_b 3"
def fn(g: Gaussians["batch_a batch_b"]):
    ...

Features

Shape Inference

A @tensorbox class will automatically infer its batch shape:

@tensorbox
class Camera:
    intrinsics: Float[Tensor, "3 3"]
    extrinsics: Float[Tensor, "4 4"]

cameras = Camera(
    torch.zeros((512, 4, 3, 3), dtype=torch.float32),
    torch.zeros((512, 4, 4, 4), dtype=torch.float32),
)

cameras.shape  # (512, 4)

Nested Tensorboxes

You can define and use nested @tensorbox classes as follows:

@tensorbox
class Leaf:
    rgb: Float[Tensor, "3"]
    scale: Float[Tensor, ""]

@tensorbox
class Tree:
    pair: Leaf["2"]

def fn(tree: Tree["*batch"]):
    # tree.pair.rgb has shape (*batch, 2, 3)
    ...

Interaction with PyTorch

@tensorbox classes can be used directly with the following torch functions:

• torch.cat
• torch.stack

Note that dim arguments are always specified relative to the @tensorbox class's batch shape.

Comparison with TensorDict

tensorbox is very similar to TensorDict, but has a few key differences:

• It's compatible with jaxtyping annotations.
• It's not as feature-complete.
• When creating a tensorbox class instance, you don't have to specify the batch shape—it's automatically inferred.