EulerFlow implementation

[jang-1996]

Hi, thank you for releasing this repository.

I am running EulerFlow on the Argoverse 2 Sensor Dataset and experiencing extremely slow validation runtime.

I run validation with the following command:

python test_pl.py ./configs/eulerflow/argo/noncausal_minibatched/val.py --gpus 1

For my setup:

Dataset: Argoverse 2 Sensor Dataset

GPU: Single RTX 5090

However, it takes approximately 14 hours to process a single validation log, which seems unusually slow even considering the cost of Eulerian ODE-based scene flow.

Could you please clarify:

Is this runtime expected for EulerFlow on Argoverse 2?

Are there known validation-time bottlenecks (e.g., ODE integration steps, or minibatch configuration)?

Are there recommended configuration changes or flags (e.g., fewer ODE steps, different batching, or mixed precision) to significantly reduce validation time?

If helpful, I can share more details about my environment or profiling results.

Thank you for your help.

[aartykov]

Hello. In the paper, they mentioned that it takes 24h for a single Argoverse2 sequence to optimize. So 14h seems to be normal. Good luck!

[jang-1996]

Thank you for your reply.

I have another question!
In "/configs/eulerflow/argo/noncausal_minibatched/test.py," SEQUENCE_LENGTH is set to 150.

I understand the paper uses all frames. Is there a reason you set it to 150?

[aartykov]

Hello! Did you first run train_pl.py to optimize the scene? How long did it take for a single Argoverse scene?

[jang-1996]

Since EulerFlow is based on test-time optimization and explicitly states that no pretraining is required, I started by executing test_pl.py first. However, I wanted to double-check if train_pl.py needs to be run beforehand. Could you please clarify the correct execution order?

Furthermore, even when I run train.py located in the config folder, the subsequence_length appears to be set to 150.

[aartykov]

Yes, but the test_pl.py file calls only the validate function, which does not optimize the MLP weights. I do not know the correct way to run the code, since I am not the author. Currently, trying to understand the code and run the method with my custom synthetic point cloud data.

[jang-1996]

Thank you for your hard work!
Running test_pl.py calls the inference_forward_single function of the WholeBatchOptimizationLoop class (models/whole_batch_optimization/whole_batch_optim_loop.py), where optimization seems to be executed.

In addition, while AV2 validation sequences typically consist of 150–190 frames, the current code seems to process only 150 consecutive frames per sequence. Since the paper reports using all frames, I would appreciate clarification on how it aligns with the experimental setup described in the paper.

[aartykov]

Do not they call "EulerFlowOptimizationLoop" class instead?

[jang-1996]

EulerFlowOptimizationLoop inherits from MiniBatchOptimizationLoop, which itself extends WholeBatchOptimizationLoop. During execution, the optimization logic defined in EulerFlowOptimizationLoop is invoked within the control flow of WholeBatchOptimizationLoop!