Add ability to use different Objectdetection models than FasterRCNN

[julled]

🚀 Feature

It would be good to be able to use more up to date models than FasterRCNN in raster-vision.
Torchvision offers some possible dropin replacements / alternatives like RetinaNet or SSD based detectors.

Motivation

More up to date models promise better performance at lower GPU needs.

Pitch

The object_detection_learner in raster-vision should be able to accept a configuration parameter which allows a different model to be used, as written above.

Alternatives

The user could do this manually in the code, but this is not very scaleable and unclear how many manual changes are needed.

Additional context

[AdeelH]

I agree that it would be useful to support other TorchVision object detection models. Adding something like an architecture field to ObjectDetectionModelConfig and then modifying ObjectDetectionModelConfig.build_default_model() might be the way to go.

That said, one way to do this without any changes to RV is to use the workaround described here:

raster-vision/rastervision_pytorch_backend/rastervision/pytorch_backend/examples/object_detection/cowc_potsdam.py

Lines 143 to 180 in 3eb3ff1

	if external_model:
	"""This demonstrates how to use an external model for object detection,
	but to successfully use this functionality with different settings, the
	following things should be kept in mind:
	(1) Torchvision does not expose its object detection models via
	torch hub (https://github.com/pytorch/vision/issues/1945). So, to
	use those, you will need to fork the torchvision repo and manually
	add those models or corresponding factory functions to hubconf.py.
	Example: github.com/AdeelH/vision/blob/det_hubconf_0.10/hubconf.py.
	Further, you should ensure that the branch of the fork is the same
	version as the version in Raster Vision's Docker image; or, if
	using outside Docker, it should match the version of the local
	torchvision installation.
	(2) The external model should behave exactly like torchvision
	object detection models. This includes, but might not be limited
	to:
	- Accepting targets as dicts with keys: 'boxes' and 'labels'.
	- Accepting 1-indexed class labels.
	- Computing losses internally and returning them in a dict
	during training.
	- Returning predictions as dicts with keys: 'boxes', 'labels',
	and 'scores'.
	"""
	model = ObjectDetectionModelConfig(
	external_def=ExternalModuleConfig(
	github_repo='AdeelH/vision:det_hubconf_0.12',
	name='ssd',
	entrypoint='ssd300_vgg16',
	force_reload=True,
	entrypoint_kwargs={
	# torchvision OD models need add an additional null class,
	# so +1 is needed here
	'num_classes': len(class_config.names) + 1,
	'pretrained': False,
	'pretrained_backbone': True
	}))

You'll need to fork the torchvision repo and create a branch with hubconf.py modified like this: https://github.com/AdeelH/vision/blob/356a78a894acf2a377353a4cdb2df52c1cea84b4/hubconf.py#L68-L72