Env package #44

zhangk1551 · 2024-04-09T00:17:48Z

Merged the torchdrivesim master into first-release-env and cleanup the map and traffic control related code. The code for the stop_sign violation is not in this PR, because it's not used in torchdriveenv now.

Since most related changes are already checked in the master, this PR is quite light-weighted.

I moved the resources folder under the torchdrivesim folder, because when I packaged the lib via pip install "torchdrivesim @ git+https://github.com/inverted-ai/torchdrivesim.git@first-release-env", it's a workable way to include the resources data to the library. Let me know if there is a better way to package the resources.

After this PR is merged, I will change the git submodule in the torchdriveenv as master. Once the torchdrivesim master is released to pip, it can be installed without a specific branch in torchdriveenv.

…ivesim into modification_for_env

zhangk1551 · 2024-04-09T00:49:23Z

torchdrivesim/behavior/iai.py

@@ -30,19 +32,84 @@ def iai_initialize(location, agent_count, center=(0, 0), traffic_light_state_his
    return agent_attributes, agent_states, response.recurrent_states


-def iai_drive(location: str, agent_states: Tensor, agent_attributes: Tensor, recurrent_states: List,
-              traffic_lights_states: Optional[dict] = None):
+def iai_conditional_initialize(location, agent_count, agent_attributes=None, agent_states=None, recurrent_states=None, center=(0, 0), traffic_light_state_history=None):


It might also be used by code outside of torchdriveenv if needed, so I keep it here.

That is specifically what I would like to avoid. I don't want this version to proliferate, so let's move it to torchdriveenv. If it stays here, it will be harder to update later.

zhangk1551 · 2024-04-09T00:52:14Z

torchdrivesim/behavior/iai.py

@@ -62,12 +129,15 @@ class IAIWrapper(NPCWrapper):
    """
    def __init__(self, simulator: SimulatorInterface, npc_mask: TensorPerAgentType,
                 recurrent_states: List[List], locations: List[str], rear_axis_offset: Optional[Tensor] = None,
+                 car_sequences: Optional[Dict[int, List[List[float]]]] = None,


It's used by torchdriveenv to set parked car, and can also be used to load preset logs.

Potentially, it can be replaced by ReplayWrapper with careful design, but currently, let's leave it here.

I can accept that, but can you at least make it replay_states (BxAxTxSt float) and replay_mask (BxA or BxAxT bool) so that we use tensors consistently, rather than resorting to ad-hoc formatted dictionaries?

adscib

Almost there. Do you still need that resources folder at all? We now install lanelet2 from pip, and the .osm files and meshes are available using the new map files format if needed.

adscib · 2024-04-09T19:34:53Z

pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
    "scipy",
    "imageio",
    "torch>=1.10.1",
-    "invertedai",
+    "invertedai==0.0.16",


Suggested change

"invertedai==0.0.16",

"invertedai",

I think this will work with older versions of invertedai now as well. We can set this constraint in torchdriveenv instead.

adscib · 2024-04-09T19:37:43Z

torchdrivesim/behavior/common.py

+from invertedai.error import InvertedAIError
+
+
+class InitializationFailedError(InvertedAIError):


If we inherit from InvertedAIError, I would move this to torchdrivesim.behavior.iai.

I just found this class was not only used in the iai related cases (

torchdrivesim/examples/check_map_alignment.py

Line 15 in c90c51b

from torchdrivesim.behavior.common import InitializationFailedError

,

torchdrivesim/torchdrivesim/behavior/heuristic.py

Line 6 in c90c51b

from torchdrivesim.behavior.common import InitializationFailedError

). Let me restore the change.

adscib · 2024-04-09T19:38:56Z

torchdrivesim/behavior/iai.py

@@ -30,19 +32,84 @@ def iai_initialize(location, agent_count, center=(0, 0), traffic_light_state_his
    return agent_attributes, agent_states, response.recurrent_states


-def iai_drive(location: str, agent_states: Tensor, agent_attributes: Tensor, recurrent_states: List,
-              traffic_lights_states: Optional[dict] = None):
+def iai_conditional_initialize(location, agent_count, agent_attributes=None, agent_states=None, recurrent_states=None, center=(0, 0), traffic_light_state_history=None):


That is specifically what I would like to avoid. I don't want this version to proliferate, so let's move it to torchdriveenv. If it stays here, it will be harder to update later.

adscib · 2024-04-09T19:40:26Z

torchdrivesim/behavior/iai.py

+                agent_attribute_list = response.agent_attributes + outside_agent_attributes
+                agent_state_list = response.agent_states + outside_agent_states
+                recurrent_state_list = response.recurrent_states + outside_recurrent_states
+            except Exception:


This is overly broad and will capture things like keyboard interrupts and out of memory exceptions, which we don't want. You probably want to catch InvertedAIError and maybe a few others?

adscib · 2024-04-09T19:42:08Z

torchdrivesim/behavior/iai.py

+    return agent_attributes, agent_states, recurrent_state_list
+
+
+def iai_drive(location: str, agent_states: Tensor, agent_attributes: Tensor, recurrent_states: List, traffic_lights_states: Dict = None):


Would you mind fixing the type signature while you're editing this function? It should specify the list element type for recurrent_state, make traffic_lights_states Optional, and specify return type.

adscib · 2024-04-09T19:42:35Z

torchdrivesim/behavior/iai.py

+    except Exception as e:
+        raise e


This doesn't do anything and can be deleted.

adscib · 2024-04-09T19:47:41Z

torchdrivesim/behavior/iai.py

@@ -62,12 +129,15 @@ class IAIWrapper(NPCWrapper):
    """
    def __init__(self, simulator: SimulatorInterface, npc_mask: TensorPerAgentType,
                 recurrent_states: List[List], locations: List[str], rear_axis_offset: Optional[Tensor] = None,
+                 car_sequences: Optional[Dict[int, List[List[float]]]] = None,


I can accept that, but can you at least make it replay_states (BxAxTxSt float) and replay_mask (BxA or BxAxT bool) so that we use tensors consistently, rather than resorting to ad-hoc formatted dictionaries?

adscib · 2024-04-09T19:47:52Z

torchdrivesim/behavior/iai.py

                 traffic_light_controller: Optional[TrafficLightController] = None, traffic_light_ids: Optional[List[int]] = None):
        super().__init__(simulator=simulator, npc_mask=npc_mask)

        self._locations = locations
        self._npc_predictions = None
        self._recurrent_states = recurrent_states
+        self._car_sequences = car_sequences
+        self._iai_timestep = 0


Suggested change

self._iai_timestep = 0

self._replay_timestep = 0

Ruishenl · 2024-04-11T23:10:21Z

Probably not related to this PR, but i think we should also add a gif from the documentation page to the README so people can have a rough idea how does everything look in general.

Also, we mentioned in the README the pypi version of this package comes with the OpenCV renderer, but we haven't made any official release to PyPI since last February. So a simple pip install will not get user the latest change from master

zhangk1551

I removed the lanelet2 dependency, but I didn't quite get the ideas about how the map files can be packaged to pip. @adscib

zhangk1551 · 2024-04-12T00:25:08Z

pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
    "scipy",
    "imageio",
    "torch>=1.10.1",
-    "invertedai",
+    "invertedai==0.0.16",


zhangk1551 · 2024-04-12T00:25:38Z

torchdrivesim/behavior/common.py

+from invertedai.error import InvertedAIError
+
+
+class InitializationFailedError(InvertedAIError):


I just found this class was not only used in the iai related cases (

torchdrivesim/examples/check_map_alignment.py

Line 15 in c90c51b

from torchdrivesim.behavior.common import InitializationFailedError

,

torchdrivesim/torchdrivesim/behavior/heuristic.py

Line 6 in c90c51b

from torchdrivesim.behavior.common import InitializationFailedError

). Let me restore the change.

zhangk1551 · 2024-04-12T00:25:48Z

torchdrivesim/behavior/iai.py

@@ -30,19 +32,84 @@ def iai_initialize(location, agent_count, center=(0, 0), traffic_light_state_his
    return agent_attributes, agent_states, response.recurrent_states


-def iai_drive(location: str, agent_states: Tensor, agent_attributes: Tensor, recurrent_states: List,
-              traffic_lights_states: Optional[dict] = None):
+def iai_conditional_initialize(location, agent_count, agent_attributes=None, agent_states=None, recurrent_states=None, center=(0, 0), traffic_light_state_history=None):


zhangk1551 · 2024-04-12T06:41:41Z

torchdrivesim/behavior/iai.py

+                agent_attribute_list = response.agent_attributes + outside_agent_attributes
+                agent_state_list = response.agent_states + outside_agent_states
+                recurrent_state_list = response.recurrent_states + outside_recurrent_states
+            except Exception:


zhangk1551 · 2024-04-12T09:57:56Z

torchdrivesim/behavior/iai.py

+    return agent_attributes, agent_states, recurrent_state_list
+
+
+def iai_drive(location: str, agent_states: Tensor, agent_attributes: Tensor, recurrent_states: List, traffic_lights_states: Dict = None):


zhangk1551 · 2024-04-12T09:58:30Z

torchdrivesim/behavior/iai.py

+    except Exception as e:
+        raise e


zhangk1551 · 2024-04-12T10:02:05Z

torchdrivesim/behavior/iai.py

                 traffic_light_controller: Optional[TrafficLightController] = None, traffic_light_ids: Optional[List[int]] = None):
        super().__init__(simulator=simulator, npc_mask=npc_mask)

        self._locations = locations
        self._npc_predictions = None
        self._recurrent_states = recurrent_states
+        self._car_sequences = car_sequences
+        self._iai_timestep = 0


adscib

Looks good! Just a few nitpicks left.

torchdrivesim/resources/maps/carla_Town06/metadata.json

torchdrivesim/resources/maps/carla_Town07/metadata.json

adscib · 2024-04-12T22:14:13Z

torchdrivesim/behavior/iai.py

    import invertedai
    from invertedai.common import AgentState, AgentAttributes, Point
    agent_attributes = [AgentAttributes(length=at[0], width=at[1], rear_axis_offset=at[2]) for at in agent_attributes]
    agent_states = [AgentState(center=Point(x=st[0], y=st[1]), orientation=st[2], speed=st[3]) for st in agent_states]
+    seed = random.randint(1, 10000)


Why is this preferred to leaving seed as None? If you want reproducibility after setting global seed, that makes sense, but in that case do it consistently in both drive and initialize.

Yes, it's for reproducibility. It's set in the drive and conditional_initialize, because torchdriveenv doesn't use the original initialize.

torchdrivesim/behavior/iai.py

zhangk1551 and others added 30 commits November 23, 2023 15:44

Modifications for the RL env.

b0f6986

Move the gym_env and env_utils

aaa348d

Skip the agent type

1e0e7ec

Add waypoint env

6ee345e

Waypoint disappear after reaching

9fcc973

Multiple cases in one suite;scenario

200b02c

Enable car sequences

f554a53

Eable traffic lights preset and stop signs

337f2c7

Change reward function

fd84773

Remove previous env

2f6c3f6

Load background traffic

d981dae

Update

35c6820

Conditional initialization based on background

af15c6a

Fix typo

998ea67

Add traffic light state machine

ea8a771

Clean the code

a857af6

Rendering stop_sign and yield from location_info

00cfef1

Condition on preset NPC

38bf587

Remove traffic light state fallback

6850eb7

Fix the initialize calling

963a9df

Improve the experiment and add logging callback

2be549c

Add a flag for terminate_at_infraction

3d73aba

Update

86c266d

Change reward

8c91b80

Enable ego only training and validation

5da4c62

Change reward function

fbadfc7

Merge branch 'modification_for_env' of github.com:inverted-ai/torchdr…

e40d68e

…ivesim into modification_for_env

Disable updating the location_info cache in every run

f1f2e62

Use first waypoint as start pos if sample offroad

101e961

Add more metrics

c65bd53

zhangk1551 added 8 commits March 15, 2024 01:45

Update pyproject.toml

acd26b7

Move resources inside "torchdrivesim" dir

c286d5d

Update resources paths in code

335fd8d

Update pyproject.toml

427f01e

Merge branch 'master' into cleanup

6c7957c

Merge branch 'master' into first-release-env

5d89ee0

Merge map and traffic_controls

b4a5bed

Remove unused resources

196a463

zhangk1551 changed the base branch from master to first-release-env April 9, 2024 00:23

zhangk1551 changed the base branch from first-release-env to master April 9, 2024 00:24

zhangk1551 added 2 commits April 8, 2024 17:39

Remove unused code

6dbf8f3

Removed unused code

792a984

zhangk1551 commented Apr 9, 2024

View reviewed changes

zhangk1551 requested a review from adscib April 9, 2024 00:59

zhangk1551 added 3 commits April 8, 2024 18:36

Fix cuda

720bb5e

Move utils to torchdriveenv

8a6206f

Fix inconsistent device

c90c51b

zhangk1551 mentioned this pull request Apr 9, 2024

Migrate to new maps and traffic controls inverted-ai/torchdriveenv#4

Merged

adscib requested changes Apr 9, 2024

View reviewed changes

Fix the left-handed issue

94d3e3e

Address comments

abcefe9

zhangk1551 commented Apr 12, 2024

View reviewed changes

Remove extra maps

407dbe4

adscib requested changes Apr 12, 2024

View reviewed changes

zhangk1551 and others added 2 commits April 12, 2024 16:47

Fix comments

e07cdf7

Seed iai_initialize

83ce180

adscib approved these changes Apr 13, 2024

View reviewed changes

zhangk1551 merged commit e93c34e into master Apr 13, 2024
4 checks passed

zhangk1551 deleted the env-package branch April 13, 2024 00:22

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Env package #44

Env package #44

zhangk1551 commented Apr 9, 2024 •

edited

Loading

zhangk1551 Apr 9, 2024

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

zhangk1551 Apr 9, 2024

adscib Apr 9, 2024

adscib left a comment

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

adscib Apr 9, 2024

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

adscib Apr 9, 2024

adscib Apr 9, 2024

zhangk1551 Apr 12, 2024

Ruishenl commented Apr 11, 2024

zhangk1551 left a comment

zhangk1551 Apr 12, 2024

zhangk1551 Apr 12, 2024

zhangk1551 Apr 12, 2024

zhangk1551 Apr 12, 2024

zhangk1551 Apr 12, 2024

zhangk1551 Apr 12, 2024

zhangk1551 Apr 12, 2024

adscib left a comment

adscib Apr 12, 2024

zhangk1551 Apr 12, 2024

		from invertedai.error import InvertedAIError


		class InitializationFailedError(InvertedAIError):

		return agent_attributes, agent_states, recurrent_state_list


		def iai_drive(location: str, agent_states: Tensor, agent_attributes: Tensor, recurrent_states: List, traffic_lights_states: Dict = None):

Env package #44

Env package #44

Conversation

zhangk1551 commented Apr 9, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

adscib left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ruishenl commented Apr 11, 2024

zhangk1551 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

adscib left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zhangk1551 commented Apr 9, 2024 •

edited

Loading