The Autoware Foundation
Autonomy Software WG Meeting 2022/03/01 #2468
Replies: 0 comments 2 replies
-
@xmfcx @mitsudome-r I am lacking context, but I assume we are talking about this: https://gitlab.com/autowarefoundation/autoware.auto/AutowareAuto/-/blob/e718a68266aaa168558b38c08b99aa7f256d71ae/src/common/tvm_utility/test/yolo_v2_tiny/main.cpp#L273 The goal of the current tests is to ensure repeatability and determinism of the results for a given network. It also tests against regressions in other parts of the environment (modifications to the tvm_utility package, to the TVM runtime version, ...). Checking the label makes perfect sense and can become an additional test. But I don't think the current ones should be removed, unless the determinism of the results is not important. |
Beta Was this translation helpful? Give feedback.
-
|
@ambroise-arm I was actually referring to this: https://gitlab.com/autowarefoundation/autoware.auto/AutowareAuto/-/merge_requests/1303#note_860406908
I misunderstood the point made here. I misread it as outputs being different in float precision level, not vector element count difference. It was my fault sorry. |
Beta Was this translation helpful? Give feedback.
-
Administrative
Common Resources
Announcements
Autoware.Auto
Attendees
Chaired by Mitsudome
Discussion topics
Fatih (F) - since no future development will be done on Autoware.Auto, why should we care about tests? Mitsudome (M) - tests aren't just for developing (for example unit tests show that components of the code function as expected). It's important to have working tests for the release. F- Of course it's important but for the Yolo issue previous tests were written by hard coding bounding boxes and expecting the exact same output when Yolo runs. That test makes no sense. M - yes, that specific test makes no sense. Rather ignore the bounding box and check label outputs are correct. This just sounds like a poorly designed test.
Mitsudome - I will think about how to maintain the open MRs.
!1095 - Fatih (F) - the MR is almost finished but it may not be used since we are promoting the new Autoware. We may even rewrite the driver. Mitsudome (M) - until a new driver is ready, this would be the driver available to users. F - I would say they would use ROS 2 drivers since ours doesn't support regular Velodyne directly. M - we created a Velodyne driver for our use case. We will put drivers inside a separate repository in Github. Nagoya Uni. is also developing a better written driver so I shall check the current status with them.
Fatih - the Yolo MR is just the tests. Once it's merged we should create a node to run the inference ourselves. Mitsudome - the tracker expects pointcloud clusters and image detections. This can also be added to the planner although it is not integrated into the prediction step.
Action Items
Review of issues
The following issues were discussed (see comments for details):
Beta Was this translation helpful? Give feedback.
All reactions