71 gnn for 3d simulator data redesign #72

TOC6 · 2026-01-13T22:57:03Z

Initial code working. We can build torch geometric datasets, train GNNs on those datasets, and perform some analysis of the results.

…his branch: fixed variants problem in boidsAgents, added ground weights option, fixed separation weight getting multiplied twice, updated tqdm,

…imple) GNN model is created, which can be trained, and basic plotting of attention weights is provided.

…ectories working, and some minor fixes.

…working. Training seems to be working too well -- maybe some of the validation data is leaking into the training? Need to figure that out.

…es (which seems to work), run_simulator.py added as new (and better designed) entry point for the simulator, gnn_agent.py added to allow the gnn model to work with the simulator for rollouts, boid_agents.py added to allow for different kinds of agent classes to be used with the simulator, various other fixes and bugs added.

…ode off into a separate file.

… along the way.

…ta file. Some formatting fixes as well.

…ata directory to better separate the original data from the training results and make it easier to have multiplet types of input features and models.

… show.) - Added abstract class in simulator_agents_abstract.py - Clipped the actions to run trajectories so they don't go beyond the box boundaries.

… 5. Need more tests. Also, pytest code ran successfully.

…ter that train uses to create the model rather than the model being hard coded in train_3DGNN().

…s if running remote tests. Also changed where those folders were placed.

…. This commit has build and train. Last one was just build and it ran successfully.

…. This commit has build and train and show. Last one was build and train, which ran successfully.

…. This commit has build, train and analyze. Last one was build, train, and show, which failed.

…ild, train and analyze all run remotely, but the test for show fails for some reason. Changed the show test to run locally only.

dimkab

Thanks for the awesome work!

dimkab · 2026-01-14T16:57:35Z

collab_env/sim/gymnasium_env/envs/boidsWorldSimple.py

+        """
+        TOC -- 010226 8:52PM
+        Why are these torch.inf instead of np.inf?
+        """


have you tried to switch to np.inf and rerun tests?

All tests pass with np.inf, so I changed that.

dimkab · 2026-01-14T17:05:31Z

collab_env/sim/boids/run_simulator.py

we have two run_simulator()'s now (this one and the old run_boids_simulator.py) and this is confusing. Since you have run_simulator_main here with boid_agents, is it supposed to replicate the functionality from run_boids_simulator.py? if so, I'd much prefer if there was only one simulator...

run_simulator() is more general and doesn't always use boids, so it makes sense to keep that name. I removed run_boids_simulator.py.

dimkab · 2026-01-14T17:26:17Z

collab_env/gnn/gnn_3D/build_dataset.py

+
+def convert_dataframe_to_node_features(
+    df: pd.DataFrame, columns=None, time_window_length=1
+):


docstring/type hint for return value

Added docstrings and type hints for all the new code (I think) -- hopefully, they are accurate and specific enough. Didn't have time to go back to the simulator code.

dimkab · 2026-01-14T17:27:07Z

collab_env/gnn/gnn_3D/build_dataset.py

+):
+    """
+    Args:
+     df (pd.DataFrame): the dataframe containing a description of the agents in the environment


expected format of the df?

dimkab · 2026-01-14T17:41:23Z

collab_env/gnn/gnn_3D/build_dataset.py

+    node_feature_columns=None,
+    label_type: Optional[str] = None,
+    time_window_length: int = 1,
+):


type hint / docstring for return value

dimkab · 2026-01-14T21:48:34Z

collab_env/gnn/gnn_3D/train_3DGNN.py

+    )
+
+
+def load_dataset(directory: str):


(optional) parameters: seed, batch_size (is >1 supported?), train_size_ratio

I believe batch_size>1 will work (it runs) but I have to rethink how I am processing the training results since torch-geometric combines the graphs into a single graph.

I added support in process_results() to handle results generated with batches. I also added train_size_ratio and seed parameters.

dimkab · 2026-01-14T21:49:12Z

collab_env/gnn/gnn_3D/train_3DGNN.py

+    )
+
+    train_loader = DataLoader(dataset=train_dataset, batch_size=1, shuffle=True)
+    val_loader = DataLoader(dataset=val_dataset, batch_size=1, shuffle=True)


for validation is better to have shuffle=False to compare between epochs

dimkab · 2026-01-14T21:59:54Z

collab_env/gnn/gnn_3D/train_3DGNN.py

+                if not stored_init_pos:
+                    stored_init_pos = True
+                    # print('x shape: ', graph.x.shape)
+                    """


is this comment still relevant? if anything, the assumption about attention weights being returned does create dependency to the model type.

Removed the stored init pos stuff.

dimkab · 2026-01-14T22:04:37Z

tests/gnn3D_postion_labels_agent_config.yaml

is this file used somewhere?

Nope, removed it.

dimkab · 2026-01-14T22:06:11Z

tests/test_gnn3D_position_labels.py

+    #
+    # Train
+    #
+    result = subprocess.run(


is there any way to have everything here run without spawning a new process? will help debugging the thing...

actually never mind that, debugging is not an issue

Removed all those already. :)

dimkab · 2026-01-15T18:17:43Z

tests/test_gnn3D_position_labels.py

+                scale_positions=1500,
+            )
+
+            if remote_test:


this branch is never executed

I excluded the show_trajectories part of the test for now because I couldn't get it to work. If we add this back it, we will need the remote test check to see if we want to remove the folders.

…a numpy array to a scalar. Got rid of a bunch of unnecessary comments. Added type hints everywhere. Added a notebook to run the 3DGNN code (animating weights doesn't work yet in the notebook, though). Removed the gnn3D_postion_labels_agent_config.yaml files and the run_boids_simulator file.

…. support in process_results() for batched data. Removed now redundant run_boids_simulator.py.

…_main() instead of defunct run_boids_simulator(). Updated README.md to say run_simulator where it used to say run_boids_simulator.

… run_boids_simulator().

… can get past the CI test. Removed a couple of unused files from the simulator folder.

… wouldn't have conflicts if the default config.yaml file changed. The notebook requires like 5 episodes, while the default was just 1 episode.

TOC6 added 15 commits December 23, 2025 19:00

made changes to get important changes from branch 56-simulator into t…

68db74c

…his branch: fixed variants problem in boidsAgents, added ground weights option, fixed separation weight getting multiplied twice, updated tqdm,

initial 3D GNN code included. A dataset can be built, a sample (and s…

94c6184

…imple) GNN model is created, which can be trained, and basic plotting of attention weights is provided.

3D GNN code with option for MLP and no convolutional layer, show traj…

acbb96e

…ectories working, and some minor fixes.

Matched input data file names to result file names. Trajectories are …

199b7ad

…working. Training seems to be working too well -- maybe some of the validation data is leaking into the training? Need to figure that out.

backed out changed to run_boids_simulator.py since we split the new c…

0722e86

…ode off into a separate file.

Rollout code in anaylze_results.py was added. Fixed a bunch of things…

dd54263

… along the way.

Rollout code gets the node feature columns through the dataset metada…

f9291b2

…ta file. Some formatting fixes as well.

Now allows for symbolic link in raw directory to point to the sim3d d…

8dd4f41

…ata directory to better separate the original data from the training results and make it easier to have multiplet types of input features and models.

- First test working for gnn3D. (Need to fix it so visualizer doesn't…

d94600c

… show.) - Added abstract class in simulator_agents_abstract.py - Clipped the actions to run trajectories so they don't go beyond the box boundaries.

Removed some print statements and fixed some comments.

d9f63fb

Added support for time windows. Tested for windows of length 1,2, and…

66085a2

… 5. Need more tests. Also, pytest code ran successfully.

Made train more generic by including a model creation function parame…

216818f

…ter that train uses to create the model rather than the model being hard coded in train_3DGNN().

Fixed filename include two .parquets in analyze_results.py.

301b3ce

Updated test_gnn3D_position_labels.py to remove the folders it create…

ea42e10

…s if running remote tests. Also changed where those folders were placed.

TOC6 linked an issue Jan 13, 2026 that may be closed by this pull request

GNN for 3D Simulator Data -- Redesign #71

Open

TOC6 added 6 commits January 13, 2026 18:11

Removed matplotlib("TkAgg") thing so we can do remote tests.

07f4a47

Removed matplotlib("TkAgg") thing so we can do remote tests.

466ad50

testing the test_gnn3D_position_labels.py code on CI by removing bits…

b65e57b

…. This commit has build and train. Last one was just build and it ran successfully.

testing the test_gnn3D_position_labels.py code on CI by removing bits…

770b4f5

…. This commit has build and train and show. Last one was build and train, which ran successfully.

testing the test_gnn3D_position_labels.py code on CI by removing bits…

ae906ff

…. This commit has build, train and analyze. Last one was build, train, and show, which failed.

Tested the test_gnn3D_position_labels.py code on CI. The tests for bu…

96ebcde

…ild, train and analyze all run remotely, but the test for show fails for some reason. Changed the show test to run locally only.

TOC6 requested a review from dimkab January 14, 2026 02:07

dimkab requested changes Jan 14, 2026

View reviewed changes

dimkab reviewed Jan 15, 2026

View reviewed changes

TOC6 added 5 commits January 15, 2026 14:23

Fix the saved model name in test_gnn3D_position_labels.py.

4d4bc69

Added batch size, seed, train_size_ratio parameters in train_3DGNN.py…

c4640ea

…. support in process_results() for batched data. Removed now redundant run_boids_simulator.py.

Fixed a type hint in save_predictions()

9f095e8

Finished adding the additional parameters in train_3DGNN.py.

872ed8f

TOC6 added 9 commits January 15, 2026 20:32

Removed show_agents_from_file.py. Changed tests to call run_simulator…

8422198

…_main() instead of defunct run_boids_simulator(). Updated README.md to say run_simulator where it used to say run_boids_simulator.

Changed one last test to call run_simulator_main() instead of defunct…

88e4498

… run_boids_simulator().

Sample notebook added to docs/gnn/gnn3D.

fb7a5db

Fixed the lint errors in the sample notebook.

2813dd2

Removed the simulator windows from sample.ipynb.

85eb44b

Removed the kernelspec from the metadata in sample.ipynb to see if we…

ed22511

… can get past the CI test. Removed a couple of unused files from the simulator folder.

Needed a config.yaml file specific to the notebook sample.ipynb so we…

bb2eea1

… wouldn't have conflicts if the default config.yaml file changed. The notebook requires like 5 episodes, while the default was just 1 episode.

Cleared the output from sample.ipynb.

7089900

Fix the case on the model file.

50d1f3e

71 gnn for 3d simulator data redesign #72

Are you sure you want to change the base?

71 gnn for 3d simulator data redesign #72

Uh oh!

Conversation

TOC6 commented Jan 13, 2026

Uh oh!

dimkab left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

TOC6 Jan 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

TOC6 Jan 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

TOC6 Jan 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

TOC6 Jan 14, 2026 •

edited

Loading

TOC6 Jan 14, 2026 •

edited

Loading

TOC6 Jan 16, 2026 •

edited

Loading