Visualization

This module provides tools to generate useful visualizations.

Installation

Since this module is optional, we need to install it before using it:

pip install ninetoothed[visualization]

Visualizing a Tensor

visualize can be used to visualize a single tensor:

ninetoothed.visualization.visualize(tensor, color=None, save_path=None)

Visualize a tensor as a structured grid representation.

Parameters:

• tensor – The tensor to be visualized.

• color – The color to be used for visualization.

• save_path – The path where the visualization should be saved.

Basic Usage

If we just want to temporarily visualize a tensor, the simplest way is to pass the tensor to visualize and provide a save path. Then, visualize will save the generated image at the specified location.

x = Tensor(shape=(4, 8))
visualize(x, save_path="x.png")

This method can also be used to temporarily visualize multiple tensors. You just need to pass the corresponding tensors and save paths to visualize.

x = Tensor(shape=(4, 8))
visualize(x, save_path="x.png")

y = Tensor(shape=(8, 4))
visualize(y, save_path="y.png")

z = Tensor(shape=(4, 4))
visualize(z, save_path="z.png")

Specifying Colors

When using the above method, tensors are assigned default colors. To customize colors, we can use the color parameter, following Matplotlib’s color formats.

x = Tensor(shape=(4, 8))
visualize(x, color="orange", save_path="x.png")

Visualizing an Arrangement

visualize_arrangement can be used to visualize an arrangement:

ninetoothed.visualization.visualize_arrangement(arrangement, tensors)

Visualize the arrangement of the tensors.

Parameters:

• arrangement – The arrangement of the tensors.

• tensors – The tensors.

Note: Currently, this API must be run in a CUDA-available environment.

This API requires arrangement and tensors as parameters, similar to ninetoothed.make(). However, there is a key difference: neither arrangement nor tensors can contain symbols; they can only contain concrete values.

As demonstrated below, we can use functools.partial to pass concrete arguments to the _arrangement function, constructing a symbol-free arrangement. In addition, we can create a tensor with a concrete shape using the shape parameter of Tensor.

def _arrangement(tensor, tile_shape):
    return (tensor.tile(tile_shape),)


arrangement = functools.partial(_arrangement, tile_shape=(2, 2))
tensors = (Tensor(shape=(5, 5)),)

visualize_arrangement(arrangement, tensors)