Search Issue Tracker

Fixed

Fixed in 2.3.0

Votes

0

Found in [Package]

1.2.0

Issue ID

BAR-8

Regression

No

[Barracuda]Output tensor channels are swapped and values are much higher using Pose Detector ONNX model (channel swap part - fixed)

Package: Barracuda

-

Reproduction steps:
1. Open the "UnityProject" from the latest "default" branch.
2. Make sure to run the generate-all script to update the models.
2. Open the "PoseDetectorTests.cs" script and remove the comment lines (/* and */).
3. Open the Test Runner window and run "PoseDestectorTests".

Actual result:
Received shape (1, 1, 1, 1, 1, 1, 1, 896) expected shape (1, 1, 1, 1, 1, 896, 1, 1).
Received shape (1, 1, 1, 1, 1, 1, 8, 896) expected shape (1, 1, 1, 1, 1, 896, 1, 8).
Average error over values exceeding delta threshold
Received -1935.363 -1303.264 -2113.783...
Expected -30.31235 -22.73796 -74.91953...

(The tensor shape issue is fixed. Outputs still don't match.)

Notes:
- Couldn't test older major Barracuda versions, model is too difficult for them.
- The model is also attached to the case (posedetector_full.onnx).
- Original issue: [https://github.com/Unity-Technologies/barracuda-release/issues/91]

Add comment

Log in to post comment

All about bugs

View bugs we have successfully reproduced, and vote for the bugs you want to see fixed most urgently.