Update task summary #21067
Update task summary #21067
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
stevhliu
force-pushed
the
update-task-summary-2
branch
from
27db7fa
to
d1cbf2f
Compare
stevhliu
force-pushed
the
update-task-summary-2
branch
from
d1cbf2f
to
16cf536
Compare
|
Ok, I'm finally finished with the first draft (took a bit longer to learn some models I wasn't familiar with)! I'd appreciate a general review of the scope of this page to make sure we're aligned (ie, are some sections too in-depth, are some not explained well enough?). Thanks in advance @sgugger @MKhalusova ! 🥹 Afterward, I'll ping one of our audio and computer vision experts for a more in-depth review of those sections 🙂 |
There was a problem hiding this comment.
This is a massive effort and will likely become a super useful doc once it's merged! Thank you for writing it! I feel like we should set expectations for the reader expertise level for all the sections and provide links to resources where they can learn the basics (just like the link to the course). Most sections require some familiarity with the subject, and it is common for folks to have expertise in some modality but not in all of them.
|
Thanks for the feedback, I added some images to go along with the text! @NielsRogge, would you mind reviewing the computer vision section? This guide is a high-level overview, and the goal is to help users understand how a certain task is solved by a model. Please feel free to let me know if it's too detailed, not detailed enough, or if I got something wrong! Also, if you know of a good beginner's resource for computer vision we can link to, that'd be great as well to set expectations for the reader. Thanks! 👍 @sanchit-gandhi, if you could do the same with the audio section, that'd be awesome. Thank you! 👍 |
There was a problem hiding this comment.
Very cool! Would maybe just go one level higher on the pre-training details for W2V2 (as these are pretty conceptually difficult 😅)
Also think it's fine to expand a bit on how the classification heads work. This is the key point here so maybe doubling down and making sure this is very clear and precise in how we describe it.
Also our convention has been to write the model as Wav2Vec2, which differs from facebook's wav2vec 2.0 and the wav2vec2 naming used here! Perhaps we could stay consistent with our current docs and update to Wav2Vec2?
* first draft of audio section * make style * first draft of computer vision section * add convnext and encoder tasks * finish up nlp tasks * minor edits * add arch images, more edits * fix image links * apply sanchit feedback * model naming convention * apply niels vit feedback * replace detr for segmentation with mask2former * apply feedback * apply feedback
* first draft of audio section * make style * first draft of computer vision section * add convnext and encoder tasks * finish up nlp tasks * minor edits * add arch images, more edits * fix image links * apply sanchit feedback * model naming convention * apply niels vit feedback * replace detr for segmentation with mask2former * apply feedback * apply feedback
* first draft of audio section * make style * first draft of computer vision section * add convnext and encoder tasks * finish up nlp tasks * minor edits * add arch images, more edits * fix image links * apply sanchit feedback * model naming convention * apply niels vit feedback * replace detr for segmentation with mask2former * apply feedback * apply feedback
This is the second part of updating the task summary to be more conceptual. After a brief introduction and background to the tasks Transformers can solve in part 1, this PR is a bit more advanced and digs deeper into explaining how Transformer solves these tasks.
To-do: