Conversation
|
A couple questions:
And, since the proposal is for this to ship as part of PV, two requests:
|
No, it shouldn't need updates over time unless we happen to change something about the way we do RKNN.
Yes, we'll need a different version for Rubik, as the Rubik Pi uses a different format for models.
Maybe? It depends on what you want me to test.
Hmm, I hadn't thought about that. I'll explain it under the OD section, and I think that should cover it? |
It would be good to confirm that the model is working against one or more test images, but I'm not sure if that can only be done on an OrangePi, or if it can be run as part of the CI tests on GitHub. |
Unfortunately, it can only be tested on opi unless we want to take the trouble of installing all the RKNN-toolkit libraries in CI. I don't think that's terribly feasible. |
samfreund
force-pushed
the
new-default-rknn
branch
from
06afdcc to
150da75
Compare
That's what I was worried about. So, we rely on the scream test to know if it's not working. :( |
Essentially, yes 😢 I will likely test these models in the test matrix and before we cut a release, but yea. |
Gold856
left a comment
Gold856
left a comment
There was a problem hiding this comment.
LGTM besides the concerns of not being able to test this
Gold856
left a comment
Gold856
left a comment
There was a problem hiding this comment.
Actually, we should document this.
samfreund
force-pushed
the
new-default-rknn
branch
from
150da75 to
b941778
Compare
samfreund
added
enhancement
This reverts commit 7531238.
This reverts commit 7531238. ## Description The COCO dataset contains images that use the NC and/or ND variants of the CC license, and distributing a model based on that dataset is most likely a violation of licenses. Additionally, the model is licensed under AGPL, which might be a concern for PhotonVision, and at a minimum, there's no license file bundled with the model right now. ## Meta Merge checklist: - [x] Pull Request title is [short, imperative summary](https://cbea.ms/git-commit/) of proposed changes - [x] The description documents the _what_ and _why_ - [x] If this PR changes behavior or adds a feature, user documentation is updated - [ ] If this PR touches photon-serde, all messages have been regenerated and hashes have not changed unexpectedly - [ ] If this PR touches configuration, this is backwards compatible with settings back to v2024.3.1 - [ ] If this PR touches pipeline settings or anything related to data exchange, the frontend typing is updated - [ ] If this PR addresses a bug, a regression test for it is added
## Description See #2026 for the previous iteration of this PR. This adds the RKNN model trained on the COCO dataset as one of the models shipped with PV. This model is fairly general, and has been trained to identify a number of objects, including people, animals, cars, and more. This model is meant for teams to test object detection, particularly for teams who might not have access to the game elements that our other models are trained on. It additionally acknowledges Ultralytics for the model, and includes the AGPL copyleft license. ## Meta Merge checklist: - [x] Pull Request title is [short, imperative summary](https://cbea.ms/git-commit/) of proposed changes - [x] The description documents the _what_ and _why_ - [x] If this PR changes behavior or adds a feature, user documentation is updated - [x] If this PR touches photon-serde, all messages have been regenerated and hashes have not changed unexpectedly - [x] If this PR touches configuration, this is backwards compatible with settings back to v2024.3.1 - [x] If this PR touches pipeline settings or anything related to data exchange, the frontend typing is updated - [x] If this PR addresses a bug, a regression test for it is added
3 participants
Description
This adds the RKNN model trained on the COCO dataset as one of the models shipped with PV. This model is fairly general, and has been trained to identify a number of objects, including people, animals, cars, and more. This model is meant for teams to test object detection, particularly for teams who might not have access to the game elements that our other models are trained on.
Meta
Merge checklist: