Aim v2.6.0 — Docker requirement removed and Metric as x-axis

Gev Sogomonian
Jun 24 · 2 min read

Aim 2.6.0 is out! We are on a journey to democratize AI dev tools and thankful to the community for continuous feedback.

Check out the updated Aim demo at play.aimstack.io.

For this release, there are two major updates — both highly requested by the Aim community. Specifically thanks to jaekyeom, siddk and vopani for raising and pushing for those improvements. See the full list of changes here and here.

Docker requirement removed

Back when we were designing the first version of Aim, we used to be excited about packaging Aim in a docker — had spent considerable amount of time to build a small alpine-based image-size to serve the web app.

We learned during the past 6 months that docker poses major problems in getting up and running with Aim on many of the cluster environments — requiring user permissions and creating lots of overhead.

From Aim v2.5.0 docker requirement is no longer there. We have repackaged the whole web app into Aim and structured the codebase to include all the artifacts and dependencies that come with Aim UI as well.

Before and after Aim v2.5.0

Metric as an alternative x-axis

Sometimes when comparing the metrics, it’s not enough to just put them into perspective via subplots.

For instance in the quick demo below, we are looking at a Neural Machine Translation (NMT) problem where two metrics are being compared — loss andbleu. One of the main indicators of a performant model in NMT, is when higher values of bleu correlates with the lower values of loss.

Just plotting them besides each other sometimes isn’t enough to see how they correlate. In this case we use loss as the x-axis to directly see the correlation between the lossand thebleu metrics for two groups of runs ( max_k==3and max_k == 1).

Demo for metric as alternative x-axis

Using a different metric as the x-axis is a handy tool to better see the correlation between two metrics of training runs.

Note: if the metric is not increasing monotonously, its values are re-ordered and then set as the x-axis (see in case of the loss).

Learn More

Aim is on a mission to democratize AI dev tools.

We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.

Try out Aim, join the Aim community, share your feedback, open issues for new features, bugs.

And don’t forget to leave Aim a star on GitHub for support 🙌.

AimStack

A super-easy way to record, search and compare 1000s of AI training runs.