Tensor Visualizer

This is a package to visualize tensor data in notebooks.

Install

pip install tensor-visualizer

Usage

You can use this widget in a notebook to visualize tensor data. The widget supports numpy ndarrays, PyTorch and Tensorflow tensors, as well as MLX tensors.

from tensor_visualizer import TensorVisualizer

# Visualize a 4-dimensional numpy array.
import numpy as np
data = np.random.randn(5, 10, 10, 10)
widget = TensorVisualizer(data)
widget

from tensor_visualizer import TensorVisualizer

# Visualize a 4-dimensional torch tensor.
import torch
data = torch.randn((5, 10, 10, 10))
widget = TensorVisualizer(data)
widget

You can name dimensions with the names argument:

TensorVisualizer(data, names=["batch", "channel", "height", "width"])

and label indices for the dimensions with the labels argument:

TensorVisualizer(
    data,
    names=["batch", "channel", "height", "width"],
    labels=[["b1", "b2", "b3"], ["ch1", "ch2"]]
)

By default the widget infers the color scale automatically from your data. To configure the color scale, you can set the scale_domain, scale_type, and scale_scheme properties:

# Set the scale domain to [1, 100], log scale,
# and use the viridis color scheme.
w = TensorVisualizer(data, scale_domain=[1, 100], scale_type="log", scale_scheme="viridis")
# You can also set it after creating the widget
w.scale_domain = [0, 1]

If you are using inferred scales, you can access the inferred scale properties with current_scale_domain, current_scale_type, and current_scale_scheme. These properties are available only after the widget has been shown.

You may use the permute argument the re-order the tensor dimensions in the visualization. For instance, permute=[2, 0, 1, 3] shows dimension 2 first, then dimension 0, 1, and 3. The last two dimensions are used in the heatmap.

Parameters

Development

This project consists of a Svelte library for the frontend component, and Python code for the widget.

To setup for frontend development, run:

npm install
npm run dev

You'll get a development server (usually at http://localhost:5173) that hosts an demo page of the frontend component.

To build the frontend package, run:

npm run build

To build the Python package, run:

hatch build

To develop the widget, you can start a Jupyter Lab instance and load the example notebooks in the examples folder:

hatch run jupyter lab