Getting SceneOptimizer to run on macos

[johnwlambert]

CUDA not properly supported on OSX, but CUDA is desired when available on linux to accelerate deep front-ends.

[johnwlambert]

@ayushbaid I tried out this env on my mac today. Still not functional:

(gtsfm-v2) dept-143-215-34-104:gtsfm johnlambert$ python gtsfm/scene_optimizer.py 
Traceback (most recent call last):
  File "gtsfm/scene_optimizer.py", line 19, in <module>
    from gtsam import Cal3Bundler, PinholeCameraCal3Bundler, Pose3, Rot3, Unit3
  File "/Users/johnlambert/opt/miniconda3/envs/gtsfm-v2/lib/python3.8/site-packages/gtsam/__init__.py", line 2, in <module>
    from .gtsam import *
ModuleNotFoundError: No module named 'gtsam.gtsam'

[ayushbaid]

@ayushbaid I tried out this env on my mac today. Still not functional:

(gtsfm-v2) dept-143-215-34-104:gtsfm johnlambert$ python gtsfm/scene_optimizer.py 
Traceback (most recent call last):
  File "gtsfm/scene_optimizer.py", line 19, in <module>
    from gtsam import Cal3Bundler, PinholeCameraCal3Bundler, Pose3, Rot3, Unit3
  File "/Users/johnlambert/opt/miniconda3/envs/gtsfm-v2/lib/python3.8/site-packages/gtsam/__init__.py", line 2, in <module>
    from .gtsam import *
ModuleNotFoundError: No module named 'gtsam.gtsam'

This is weird. Unfortunately, I dont have a macos device to debug this issue. Maybe try uninstalling and reinstalling gtsam via pip?

[johnwlambert]

@ayushbaid I tried out this env on my mac today. Still not functional:

(gtsfm-v2) dept-143-215-34-104:gtsfm johnlambert$ python gtsfm/scene_optimizer.py 
Traceback (most recent call last):
  File "gtsfm/scene_optimizer.py", line 19, in <module>
    from gtsam import Cal3Bundler, PinholeCameraCal3Bundler, Pose3, Rot3, Unit3
  File "/Users/johnlambert/opt/miniconda3/envs/gtsfm-v2/lib/python3.8/site-packages/gtsam/__init__.py", line 2, in <module>
    from .gtsam import *
ModuleNotFoundError: No module named 'gtsam.gtsam'

This is weird. Unfortunately, I dont have a macos device to debug this issue. Maybe try uninstalling and reinstalling gtsam via pip?

Tried that, but unfortunately same issue. Also tried pip install --pre gtsam in the environment, got same issue. @ProfFan any clue what my issue might be?

[ProfFan]

I think the problem may be that the macOS build is targeting Python 3.9?

[ProfFan]

Just tried pip install with my Homebrew-installed Python 3.9. Got no problem.

[ProfFan]

Please be sure that you have no prior install of gtsam, by running pip uninstall gtsam for 3 times.

[johnwlambert]

Just tried pip install with my Homebrew-installed Python 3.9. Got no problem.

Thanks @ProfFan . I was able to get the import to work right away when I changed the version to 3.9 in the conda environment file (we have it as 3.8 there currently).

Dask has a few compatibility issues in 3.9 though on my mac, that i didn't see on python 3.7 or 3.8 though on mac or linux. See my report here: #74

I propose we use python 3.8 for now to avoid those issues.

Also, @ayushbaid @dellaert we won't be able to run python gtsfm/scene_optimizer.py until we have a GTSAM wheel with the serialization PR merged since multiprocessing is our hard-coded dask setting currently, otherwise we'll get AttributeError: 'gtsam.gtsam.SfmData' object has no attribute 'serialize'.

@ayushbaid should we make that a config setting to determine to use threads vs. processes?

[dellaert]

I'm urgently awaiting this :-) Please email me at GT account when this is ready to try for me by just checking out this branch and following README....

[dellaert]

Does that mean it works now!?

[dellaert]

I have not gotten an email :-)