In the current state of the tests, they are running on an environment with a GPU and the correct datasets/models downloaded (i.e. ycbv and ycbv models). There is still one issue with the tests relative to the training (i.e. detector and pose) : both tests are passing individually, but one runs one then the other, the second test has an issue due to torch distributed. Currently, the second test (pose) is skipped
TODO short term :
Add the test results to the repository
Solve the torch distributed issue
Use HOPE models / dataset for the tests instead of YCBV, so the download are lighter
TODO long term:
Create a dedicated test configuration for the training/evaluation
Create a short dataset to run this tests quickly (they might be quite long right now)
The test could be seen as interdependent i.e. it could be organized the following way : inference on a test dataset, then evaluation of this inference. Currently, everything is independent.
In the current state of the tests, they are running on an environment with a GPU and the correct datasets/models downloaded (i.e. ycbv and ycbv models). There is still one issue with the tests relative to the training (i.e. detector and pose) : both tests are passing individually, but one runs one then the other, the second test has an issue due to
torch distributed. Currently, the second test (pose) is skippedTODO short term :
test resultsto the repositorytorch distributedissueTODO long term: