Tensor-Spectacle: Visualizing a Sequence of Tensors

Uses manim to visulize a sequence of tensors. This repository is currently used for personal academic paper drawing purpose. If you are interested in codevelop this further, please get in touch.

demo takes a few seconds to load:

• When using the matplotlib as engine:

• Native mainim cubes (much much slower)

Installation:

pip install Tensor-Spectacle

Requirements

• LaTex installed
• Whatever other packages on which manim is based

Quick Start:

To visualize a list of tensors, you can easily create a class as shown below:

from tensorspec.scenes.tensors_distribution import TensorVisualizationScene

# make some random 4D tensors (batch, channel, h, w)
random_shapes = [(randint(1, 2), randint(3, 6), randint(3, 6), randint(3, 6)) for _ in range(3)]
tensors = [torch.rand(shape) for shape in random_shapes]
labels = [f"Random Tensor {i}" for i in range(len(tensors))]

# make scene
class Example(TensorVisualizationScene):
    def construct(self):
        super().construct(tensors=tensors, labels=labels, duration_each=0.8, duration_gap=1, engine="matplotlib")

# render
with tempconfig(
    {
        "use_opengl_renderer": True,
        "disable_caching": True,
        "write_to_movie": True,
        "resolution": "1080p",
        "renderer": "cairo",  # "opengl", "cairo"; opengl is faster but manim community at the moment has some issues
        "fps": 60,
    }
):
    scene = Example()
    scene.render()

Then run your script, and watch it happen.

Note:

At the moment, using a tensor with channel size less than 3 is recommanded.

Example

An example is at _test.py

License:

This project is licensed under the CC BY-NC License.

Acknowledgments:

• Manim Community
• PyTorch

Contribution

The initial motivation was to visualize the training dynamics of biomimetic neurons. However, I realized this could be useful in many scenarios. I will continue updating this as long as I work with neural networks. Contributions are welcome.

System

Tested on Ubuntu 20.04 (see /.github/workflows)

Citations

@misc{Tensor-Spectacle2023,
  author = {Don Yin},
  title = {Tensor-Spectacle: Visualizing a Sequence of Tensors},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub Repository},
  howpublished = {\url{https://github.com/YourUsername/Tensor-Spectacle}},
}