Comparison to Lightning Fabric and ignite

[ksachdeva]

Hi,

I must say I like this little framework.

However, Lightning has something similar called Fabric and then there is this "ignite" framework with in Pytorch organization.

How do you tnt in comparison to them? What are some of the goals and non-goals?

Since it is part of official pytorch repo is there a certain level of commitment that it will be maintained and evolve

Regards
Kapil

[daniellepintz]

Hi Kapil, thanks for your kind words and your questions!

The goal of TNT is to provide a lightweight training framework as well as standalone utilities which are agnostic to any training framework. We are able to keep the framework minimal by putting as much as possible into the standalone utilities (torchtnt/utils) rather than in the framework itself. This also allows users to benefit from the utils if they don’t wish to use the framework.

Compared to Lightning Fabric, TNT runs the training/evaluation/predict loop for the user.

Regarding level of commitment we have a dedicated team at Meta supporting this framework and evolving it to keep up with the latest advances in PyTorch and ML training.