Clean up object types before one-hot encoding

[daphne-cornelisse]

Description

This PR improves the preprocessing of road and object-type features before one-hot encoding. It fixes a bug that was occurring when running the environment on the CPU (device="cpu"), where large garbage values such as -30,000 caused issues during the one-hot encoding step.

To resolve this and improve readability, I added the entity type enums to the base environment and removed all garbage values from the tensors before the one-hot encoding step.

[SamanKazemkhani]

I think there is some redundancy here:

• An enum is an alias for an integer, so there isn't a need to map the names to their underlying representation.
• If you split ENTITY_TYPE_TO_INT into two lists- one for agent entities and one for road map entities- you can eliminate ROAD_MAP_OBJECT_TYPES and ROAD_OBJECT_TYPES because they would equal the length of the corresponding lists.

Suggested change

	self.ENTITY_TYPE_TO_INT = {
	gpudrive.EntityType._None: 0,
	gpudrive.EntityType.RoadEdge: 1,
	gpudrive.EntityType.RoadLine: 2,
	gpudrive.EntityType.RoadLane: 3,
	gpudrive.EntityType.CrossWalk: 4,
	gpudrive.EntityType.SpeedBump: 5,
	gpudrive.EntityType.StopSign: 6,
	gpudrive.EntityType.Vehicle: 7,
	gpudrive.EntityType.Pedestrian: 8,
	gpudrive.EntityType.Cyclist: 9,
	gpudrive.EntityType.Padding: 10,
	}
	self.MIN_OBJ_ENTITY_ENUM = min(list(self.ENTITY_TYPE_TO_INT.values()))
	self.MAX_OBJ_ENTITY_ENUM = max(list(self.ENTITY_TYPE_TO_INT.values()))
	self.ROAD_MAP_OBJECT_TYPES = 7 # (enums 0-6)
	self.ROAD_OBJECT_TYPES = 4 # (enums 7-10)
	self.ROAD_MAP_ENTITY_TYPES = [
	gpudrive.EntityType._None,
	gpudrive.EntityType.RoadEdge,
	gpudrive.EntityType.RoadLine,
	gpudrive.EntityType.RoadLane,
	gpudrive.EntityType.CrossWalk,
	gpudrive.EntityType.SpeedBump,
	gpudrive.EntityType.StopSign,
	self.AGENT_ENTITY_TYPES = [
	gpudrive.EntityType.Vehicle,
	gpudrive.EntityType.Pedestrian,
	gpudrive.EntityType.Cyclist,
	gpudrive.EntityType.Padding,
	]
	self.MIN_OBJ_ENTITY_ENUM = min(self.ROAD_MAP_ENTITY_TYPES + self.AGENT_ENTITY_TYPES)
	self.MAX_OBJ_ENTITY_ENUM = max(self.ROAD_MAP_ENTITY_TYPES + self.AGENT_ENTITY_TYPES)

Moreover, I'm not sure these should be instance variables.

[daphne-cornelisse]

Thanks for pointing that out, I made the changes

[daphne-cornelisse]

Question: I still need to map the enums to integers. How would you suggest to do this?

Instead of the dictionary above, I could do something like:

{
     gpudrive.EntityType._None: int( gpudrive.EntityType._None),
...
}

[SamanKazemkhani]

You can use a comprehension instead (ie. {(k, int(k)) for k in self.ENTITY_TYPES}) but looking through the functions one_hot_encode_*, I don't immediately see where this mapping is necessary.

[daphne-cornelisse]

True, and it is used in the torch environment, e.g. see line 269 in env_torch.py in this PR

[SamanKazemkhani]

Suggested change

	road_types.long(), num_classes=self.ROAD_MAP_OBJECT_TYPES,
	road_types.long(), len(self.ROAD_MAP_ENTITY_TYPES),