EuroSciPy 2024

Visualizing and debugging tensors with aesthetic-tensor

Researchers and ML Engineers working hands-on with tensors frequently have to inspect and visualize the contents of tensors in order to understand what they hold. Aesthetic-tensor is a Python library that works for PyTorch and NumPy that abstracts away tensor visualizations in an intuitive and easy-to-learn API. In this tutorial, we're going to explain the core concept of the API and walk through a few practical examples of how to use the library.


Showcase of the Aesthetic-Tensor Library for visualizing and debugging tensors (Link - https://github.com/ichko/aesthetic-tensor).
- Better repr showing you the range, shape, basic stats, and problems (nans, infinities) of the tensor.
- Basic visualizations - heatmaps, histograms. GIF, etc
- Visualizing batched tensors - how to plot and integrate them in a single view

The only prior knowledge needed is for the participants to be familiar with PyTorch tensors or NumPy nd-arrays.


Abstract as a tweet:

Aesthetic-tensor, a library for visualizing and debugging tensors with dope API

Category [Community, Education, and Outreach]:

Learning and Teaching Scientific Python

Expected audience expertise: Domain:

some

Expected audience expertise: Python:

some

Public link to supporting material:

https://github.com/ichko/aesthetic-tensor

Project Homepage / Git:

https://github.com/ichko/aesthetic-tensor

The speaker’s profile picture
Iliya Zhechev

I finished masters in AI at Sofia University and currently work as an ML Engineer at MYX Robotics, doing projects related to variety of topics: Image segmentation, NLP, fault detection and more. I like to understand concepts by visualizing them interactive. My academic interests include neural cellular automatas and the emergent properties they excibit.

I am an amateur guitarist and avid shower singer.
If you meet me ask me about what is life from the point of view of information theory.