opendr-eu
Continual Transformer Encoder
New tool for efficient application of Transformer Encoders to time-series data. Note: this does not come with pre-trained models. The user must fit the learner on her own features.
I assume this PR is still ongoing, I've converted it to draft for now
It is ready for review
Thank you for updating the docs and adjusting the formatting, @ad-daniel .
Per your suggestion, I've updated the structure of the docs, splitting documentation for this module into its own file and updating index.md accordingly. Here, I've also collapsed the "action recognition" and "activity recognition" categories into one (action/activity is often used interchangeably in the literature).
Moreover, the reason for the delay to this PR was a large augmentation to the continual-inference library that enables exporting ONNX models with the accelerated "forward_step" mode. This module has been updated with that augmentation, which is enabled given the correct library version (torch >= 1.10, onnxrunetime >= 1.11, and continual-inference >= 1.0.0). Dynamic checks are made at runtime to enforce these.
There seem to be a lot of failures in tools other than this one since develop was merged (perception/activity_recognition passes). I assume this is not related to this tool, but let me know if can be of any help.
I've re-run the test a couple of times and it fails rather consistently (although in other PR targeting develop everything passes nearly every time), so I assume that since the dependency constraints here are a bit less strict, some dependencies are installed differently. One of the failures is due to this https://github.com/opendr-eu/opendr/pull/369, and as some stuff isn't installed due to the error it leads to problems. Might be worth waiting until that PR is merged
Looks like that was indeed the source
Great, thanks for handling this issue
No problem, I believe the demo and changelog entry are still missing though, no? (I assume you won't be providing a ROS interface at the moment?)
My bad. I've updated the CHANGELOG and added a demo. I will not be providing a ROS node at this time.
Thank you Lukas! I left some comments regarding some minor typos, etc.
Thanks for the thorough review, @tsampazk . I've accepted all your suggestions and fixed typos as requested. The code itself was left unchanged, so these changes should not affect anything you reviewed earlier, @ad-daniel.