traffic history improvements for imitation learning #741

sah-huawei · 2021-04-06T21:04:43Z

Changes traffic history file format from JSON to an opaque .shf format (currently a SQLite file database).

This reduces the size of history files by over 60%, speeds up stepping from the TrafficHistoryProvider to under 2ms, and fixes bugs related to issue #407 (missed and skipped samples).

This obsoletes the Traffic_history_service (and associated unit tests of it).

Traffic history database files are automatically created when a Scenario specifies them, as before, like:

gen_scenario(
    t.Scenario(
        traffic_histories=["i80-0400.yaml", "i80-0500.yaml"],
    ),
    output_dir=Path(__file__).parent,
)

However we now support a yaml "dataset spec" file as input that contains a pointer to the original dataset to be converted as well as parameters related to the conversion (examples have be added to Issue #732). This converts and imports the data into the sqlite database when scenarios are first built and obsoletes the previous conversion scripts ( tools/interaction_dataset_converter.py and ngsim_dataset_converter.py).

Along the way:

added ability to set default agent speed to the imitation learning "agent replacement" example
added ability to support different lane widths in off-road checks (as required by the NGSIM dataset, whose lanes are ~6x)
added ability to convert old JSON history files to the new format
added ability to run SMARTS in "near real-time" such that traffic can be watched in sumo-gui at correct speed

Closes #732, #407

- better smoothing of heading for very low vehicle speeds to reduce "wiggle" - added ability to set agent speed to imitation learning replacement example - added option to convert old JSON traffic histories to new sqlite .shf files - quiet netconvert's "Success" output when shifting maps - minor refactor of genhistories database creation code

JingfeiPeng · 2021-04-08T00:13:28Z

smarts/core/renderer.py

@@ -58,6 +58,7 @@ def __new__(cls):
            # disable vsync otherwise we are limited to refresh-rate of screen
            loadPrcFileData("", "sync-video false")
            loadPrcFileData("", "model-path %s" % os.getcwd())
+            loadPrcFileData("", "model-cache-dir %s/.panda3d_cache" % os.getcwd())


is this related to replaying traffic data?

Oops. No, it's not. I accidentally left that in after an unrelated test. I'll take it back out. Thanks for catching that!

(It did speed up rendering though, so we might consider adding it at some point.)

JingfeiPeng · 2021-04-08T00:28:50Z

smarts/core/sumo_road_network.py

        self._log = logging.getLogger(self.__class__.__name__)
        self._graph = graph
        self._net_file = net_file
+        self._default_lane_width = (
+            default_lane_width if default_lane_width is not None else 3.2


should we add some explanation for 3.2 here?

Yes, I will add a comment.

This is fine too but I might have just preferred a descriptive constant.

JingfeiPeng · 2021-04-08T00:48:19Z

smarts/sstudio/genscenario.py

+    genhistories_py = os.path.join(
+        os.path.dirname(os.path.realpath(__file__)), "genhistories.py"
+    )
+    for hdsr in histories_datasets:


I am wondering what does hdsr and hds stand for?...

hds -> "history data set" and hdsr -> "history data set ref" (or something like that!) :)

JingfeiPeng · 2021-04-08T00:50:02Z

cli/studio.py

@@ -146,6 +146,7 @@ def _clean(scenario):
        "social_agents/*",
        "traffic/*",
        "history_mission.pkl",


Since we are using .shf files now we can probably remove "history_mission.pkl" here

Yeah, definitely we need to eventually, but I wanted to leave it there a little bit to get rid of any existing files on the next clean.

Cool! One other thing we can remove is the ijson dependency. line 64 and 65 in setup.py and line 48 in requirements.txt. ijson was only used for imitation learning

funny, I did, but I had to add it back. It's still being used in genhistories.py to convert old json files.

JingfeiPeng

Nice change!

JingfeiPeng · 2021-04-08T01:12:18Z

where could I take a look at how NGSIM and Interaction dataset scenario folder looks now?

sah-huawei · 2021-04-08T03:02:19Z

where could I take a look at how NGSIM and Interaction dataset scenario folder looks now?

I have them locally if you want me to zip them up and send them. (I don't think we should push them here though.)
Or, you can download NGSIM from the link in Issue #407 (and INTERACTION from zbzhu99's imitation learning fork) and try out the yaml file examples I put in Issue #732. That's probably better b/c then you can see if I forgot anything in my instructions.

sah-huawei · 2021-04-08T03:27:33Z

Nice change!

Thanks. Your NGSIM jupyter notebook gave me a huge head start on genhistories.py! Thanks again for that.

Gamenot · 2021-04-10T22:03:58Z

examples/history_vehicles_replacement_for_imitation_learning.py

            observations = smarts.reset(scenario)

            dones = {agent_id: False}
-            while not dones[agent_id]:
+            while not dones.get(agent_id, False):


I think this is better practice but is there any case where the default should be necessary?

Yeah, I hit that in my testing. It happened early on and I can't clearly remember the circumstances, but I suspect it was related to my having the mission.start_time wrong initially (leading to agent_id not being started yet). So I guess, when things are correct, it's not necessary, but I added it so I wouldn't crash immediately with a key error and could debug other things first.

Gamenot · 2021-04-10T22:06:11Z

smarts/core/renderer.py

+            # TODO: the following speeds up rendering a bit... might consider it.
+            # loadPrcFileData("", "model-cache-dir %s/.panda3d_cache" % os.getcwd())


Does this speed up rendering or model loading?

oh, right, model loading. (However, as vehicle models can be loaded during the simulation, it can still affect the step time.)

Gamenot · 2021-04-10T22:10:06Z

smarts/core/sumo_road_network.py

        self._log = logging.getLogger(self.__class__.__name__)
        self._graph = graph
        self._net_file = net_file
+        self._default_lane_width = (
+            default_lane_width if default_lane_width is not None else 3.2


This is fine too but I might have just preferred a descriptive constant.

Gamenot · 2021-04-10T22:37:16Z

smarts/core/smarts.py

@@ -224,7 +226,12 @@ def _step(self, agent_actions):
        extras = dict(scores=scores)

        # 8. Advance the simulation clock.
-        self._elapsed_sim_time += dt
+        self._elapsed_sim_time = round(self._elapsed_sim_time + dt, 3)


I would say rounding here is wrong since it means that people will unable to run the simulation at any step less than 5e-4. While such a small step would not normally be done, I do not think it is on us to try limit a user from attempting a micro-scale time-step.

Good point. (I added it because the addition of dt sometimes caused floating point precision issues, like an _elapsed_sim_time equal to, say, 1.9999999 instead of 2.0.) I just changed it to never "round too much".

Gamenot · 2021-04-10T23:00:42Z

smarts/core/smarts.py

+        self._use_realtime_clock = scenario.traffic_history and (
+            not self._traffic_sim.headless or not self._envision.headless
+        )


I see why the real-time option might be useful for imitation learning but I have a few concerns. My first inclination here is that it should not be attached to the traffic history because there are other uses for a realtime clock option and this binds the option to the traffic history.

Secondly, Envision is already supposed to play back at real-time and I believe this complicates fixing Envision so I do not think Envision should be included with this option.

Otherwise this is useful to sumo-gui but there is another alternative. It is possible to send the --gui-settings-file option to sumo-gui to add step delay and breakpoints via a configuration file: https://sumo.dlr.de/docs/sumo-gui.html#configuration_files.

Re the traffic-history limitation, I agree with you. I just was hesitant to auto-set it and change the behavior for other things that might have inadvertently set headless to False but where no one watches the gui/envision.

Re Envision, there is still a bug there, but you're right that this "masks" the problem.

I will look into the --gui-settings-file option...

Using step delay via a --gui-settings-file could work assuming we know (approximately) how long our steps really take. But if these will be variable, or if we'll be improving our step time, then we may have to tweak this occasionally.

I'm willing to do this, but let's chat about it on Monday first. I do see other advantages to having a near-real-time mode (in particular, asyncrhonously running alongside ROS nodes), but there may be better ways to achieve that, so I can be persuaded to leave this out for now (and/or use the sumo-gui delay setting).

Gamenot · 2021-04-10T23:12:12Z

smarts/core/traffic_history_provider.py

-        self.start_time_offset = 0
+        self._replaced_vehicle_ids = set()
+        self._start_time_offset = 0
+        self._histories_db = None


Looks like you ordered the instance variables but then added one to put it back out of order. 🤣

hehe, yep! :)

Gamenot · 2021-04-10T23:16:36Z

smarts/core/traffic_history_provider.py

+        # Options from NGSIM and INTERACTION currently include:
+        #  1=motorcycle, 2=auto, 3=truck, 4=pedestrian/bicycle
+        # But we don't yet have glb models for 1 and 4.


I will add this as an issue.

Gamenot · 2021-04-10T23:28:42Z

smarts/core/traffic_history_provider.py

+        # But we don't yet have glb models for 1 and 4.
+        if vehicle_type == 3:
+            return "truck"
+        return "passenger"


I see this as a potential issue that we would default to a passenger vehicle when provided a pedestrian since we do not support non-road actors yet nor do we yet want to evaluate pedestrians.

Yeah, I agree. I'm adding a change to skip pedestrians for now.

... although on second thought, I think we should probably just add a pedestrian model fairly soon because, in an imitation learning scenario, it could cause problems to train from a vehicle that swerves or stops suddenly for seeming no reason (if a pedestrian that triggered this behavior is not part of the simulation).
So I guess I'm not going to skip them after all. We should just do Issue #756 relatively soon instead.

Yes, more-so I was worried about imitation learning attempting to take over a pedestrian because it thinks it is a passenger vehicle.

Ah oh. I just added something to prevent that: only passenger cars will now be included in the agent missions used by history_vehicle_replacement_for_imitation_learning.py.

Gamenot · 2021-04-10T23:37:36Z

smarts/sstudio/genhistories.py

+                float(self.column_val_in_row(row, "speed")) * self.scale,
+                self.column_val_in_row(row, "lane_id"),
+            )
+            if not any(a is not None and np.isnan(a) for a in traj_args):


I think the presence of isnan would be a value useful to warn about rather than silently skipping.

yeah, in general I agree. Unfortunately, the rolling window method used to calculate the moving averages for position_x and position_y leaves 1-window-width of NaNs at the end. This was the easiest way to ignore those. I'll add a comment to explain that.

Gamenot · 2021-04-10T23:40:21Z

smarts/sstudio/genhistories.py

+        # Try to match the NGSIM types...
+        if agent_type == "car":
+            return 2
+        elif agent_type == "truck":
+            return 3
+        elif agent_type == "pedestrian/bicycle":
+            return 4


Is motorcycle also supposed to be here?

oops, yes! good catch!

sah-huawei added 2 commits April 5, 2021 13:42

First pass at traffic history refactor to use sqlite instead of json.

34f23a8

sah-huawei requested review from Gamenot and JingfeiPeng April 6, 2021 21:04

This was linked to issues Apr 6, 2021

Traffic Histories interface bug and scalability #732

Closed

NGSIM dataset abnormal replay #407

Closed

Added copyright header.

ba69e66

This was referenced Apr 6, 2021

Traffic Histories interface bug and scalability #732

Closed

NGSIM dataset abnormal replay #407

Closed

properly take scaling (size) of road network and vehicles into account #742

Open

sah-huawei added 2 commits April 7, 2021 13:31

added ability to use watch in sumo-gui/envision in near "realtime".

2ca35e7

switch from log.warn to log.warning

24c2cd5

JingfeiPeng reviewed Apr 8, 2021

View reviewed changes

JingfeiPeng approved these changes Apr 8, 2021

View reviewed changes

updates from review

809af7a

better comment, and the black code formatter likes to mess with me.

4ced949

Gamenot approved these changes Apr 10, 2021

View reviewed changes

sah-huawei added 3 commits April 11, 2021 00:19

Fixups from review.

e1803f7

fixups from review: only replace passenger cars for imitation learning

5258dfa

removed near-real-time mode after review.

42a3518

sah-huawei merged commit f9c0f08 into develop Apr 12, 2021

sah-huawei deleted the traffic-history-sqlite branch April 12, 2021 23:38

sah-huawei mentioned this pull request Apr 22, 2021

Envision near real time #785

Merged

		# TODO: the following speeds up rendering a bit... might consider it.
		# loadPrcFileData("", "model-cache-dir %s/.panda3d_cache" % os.getcwd())

traffic history improvements for imitation learning #741

traffic history improvements for imitation learning #741

Conversation

sah-huawei commented Apr 6, 2021 • edited Loading

JingfeiPeng Apr 8, 2021 • edited Loading

Choose a reason for hiding this comment

sah-huawei Apr 8, 2021 • 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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sah-huawei Apr 8, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JingfeiPeng Apr 10, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JingfeiPeng left a comment

Choose a reason for hiding this comment

JingfeiPeng commented Apr 8, 2021

sah-huawei commented Apr 8, 2021

sah-huawei commented Apr 8, 2021

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

sah-huawei Apr 11, 2021 • 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

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

sah-huawei Apr 11, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sah-huawei commented Apr 6, 2021 •

edited

Loading

JingfeiPeng Apr 8, 2021 •

edited

Loading

sah-huawei Apr 8, 2021 •

edited

Loading

sah-huawei Apr 8, 2021 •

edited

Loading

JingfeiPeng Apr 10, 2021 •

edited

Loading

sah-huawei Apr 11, 2021 •

edited

Loading

sah-huawei Apr 11, 2021 •

edited

Loading