Open objorkman opened 2 years ago
Looks like the bug might be because deep sort sends variables to GPU devices in different sections of the code.
This might require a change to the deep sort repo.
Looks like self.gaussian_mask
is moved to a device in __init__
whereas processed_crops
is moved to a device every time run_deep_sort
is called (see here). A fix might be to move processed_crops
to the same device as self.gaussian_mask
-- I think we can do this by passing the right device when calling .cuda()
.
Screenshot:![image](https://user-images.githubusercontent.com/7193442/134785414-e906327b-a3bd-4d83-b5f6-e3de6e88dc65.png)
To reproduce, run pylot with
--flagfile=configs/tracking.conf
andtracker_type=deep_sort
set in the config file.