Upgrade torch to 1.8.1 and torchvision to 0.9.1 (feedback required)

[ad-daniel]

Notes:

• I had to change the external repository for DCNv2 from https://github.com/MatthewHowe/DCNv2 to https://github.com/jinfagang/DCNv2_latest because it was breaking for 1.8. If your tool uses it, please doublecheck if everything works fine.

Tests that work:

• engine
• utils
• perception/activity_recognition
• perception/compressive_learning
• perception/face_recognition
• perception/heart_anomaly_detection
• perception/multimodal_human_centric
• perception/pose_estimation
• perception/speech_recognition
• perception/skeleton_based_action_recognition
• perception/semantic_segmentation
• control/mobile_manipulation
• perception/object_detection_2d/
• simulation/human_model_generation
• perception/facial_expression_recognition/landmark_based_facial_expression_recognition
• perception/object_tracking_2d/deep_sort

Tests that fail:

• perception/object_tracking_2d/fair_mot. I'm getting:

from DCN import DCN
ModuleNotFoundError: No module named 'DCN'

is it supposed to be dcn module ? (lowercase)

• perception/object_detection_3d. I'm getting:

numba.cuda.cudadrv.error.NvvmError: Failed to compile

<unnamed> (54, 23): parse expected comma after load's type
NVVM_ERROR_COMPILATION

could be a local issue, can somebody else double check if it fails for you as well?

• perception/object_tracking_3d
  • This test was already deactivated. Still needs to be fixed.
  • The error is:

numba.cuda.cudadrv.error.NvvmError: Failed to compile

<unnamed> (54, 23): parse expected comma after load's type
NVVM_ERROR_COMPILATION

Missing test :

• perception/panoptic_segmentation/efficient_ps
  • the test doesn't appear to be in the list although a test is available
  • since it's added as a submodule, the change of the version has to occur on the original repo side. @vniclas if it's not a good idea to change the dependency on that end, is it possible at least to modify the current dependency from

python-dependencies=torch==1.7.1
python=torchvision==0.8.2

to

python-dependencies=torch>=1.7.1
python=torchvision>=0.8.2

?

[vniclas]

@ad-daniel We cannot run the tests for perception/panoptic_segmentation/efficient_ps in the GitHub actions since it requires a GPU, please see the discussion in PR #96. Hence, we (approved by @stefaniapedrazzi) decided to skip it in the workflow file. Generally, it is better to run all tests also locally when suggesting a global update to the toolkit.

Upgrading to PyTorch 1.8.1 is not possible for perception/panoptic_segmentation/efficient_ps as the base repo modifies code from an earlier release of mmdetection, supporting neither a CPU-only installation nor torch==1.8.1 .

[ad-daniel]

I see, that'd be a problem. It seems that the doc showed on monday wasn't complete, detectron2 is available for torch 1.7 here (expand "read more") so we might not need to do the upgrade after all. I'm waiting for Amir's response.

[ad-daniel]

It doesn't appear that it's required to upgrade to 1.8 just yet, so for now I'm closing the PR.