Aim-based experiment tracker for spaCy

[SGevorg]

Hi all, excited to share with you aim-spaCy! 🎉

It's an Aim plugin for spaCy that automatically logs the spaCy metrics and displaCy visualizations as images.
Both on training and evaluation.

Aim has a powerful Images Explorer that allows to compare 1000s of tracked images in-depth.
aim-spaCy will allow to seamlessly compare 1000s of displaCy visualizations at once with a few clicks. 🙌

A use-case example

In the gif below, we have tracked 100+ displaCy Dependency Tree and NER visualizations for one spaCy inference run.
With 2 clicks we can compare two visualizations of the same sentence at each step:

• group by record.step
• stack by the last grouping param (the step)
• [Bonus] you can compare across 100s of runs too

Basically this will allow to find where and why (thanks to displaCy magic) models misbehave / underperform.

Other perks

• Free, self-hosted and open-source
• Easy to embed in your jupyter notebook
• Compare 1000s of high-dimensional experiments (100s of metrics, images, text tracked per experiment)

web: aimstack.io/spacy
code: github.com/aimhubio/aim-spacy

Would love your feedback on aim-spaCy. What else would you like us to add to the project?

[philipvollet]

Tested around and love it! Sent you a DM via Twitter with some questions, super curious about how the project continues to develop 💙