Error on GCS evaluation

[XieShaocong33Ethan]

Hi,

The idea of lifelong crowd trajectories is super cool — I come from a more traditional pedestrian dynamics simulation background, so it’s exciting to see how your approach connects with that. Would love to exchange thoughts or even collaborate sometime if the chance comes up.

We successfully ran evaluations on the released preprocessed datasets (ETH, UCY, SDD, EDIN) without issues. However, for GCS we ran into errors during evaluation. Since the release currently provides only the raw dataset and local preprocessing instructions, we’re not sure if our locally preprocessed version matches the expected structure.

To ensure strict reproducibility with your reported results, could you provide the preprocessed GCS dataset (similar to the other datasets), or a download link to it?

If sharing a preprocessed GCS package is not feasible, could you share the exact expected output structure (folder layout and any checksums) so we can verify our locally preprocessed GCS against your reference?

• Env: Ubuntu 22.04, Python 3.10.19, PyTorch 2.2.2+cu121 (torch.version.cuda=12.1).
• Data: Downloaded raw_datasets.zip from v1.0-dataset, placed GCS under ./datasets/GCS/, and successfully ran preprocessing (outputs in ./datasets/preprocessed/gcs/).
• Error (GPU and CPU both reproduce):
  RuntimeError: to_padded_tensor: at least one constituent tensor should have non-zero numel
  Trace: nn.TransformerEncoder(...) called from CrowdES/layers.py:168, with src_key_padding_mask.

[InhwanBae]

Hi @XieShaocong33Ethan,

Thank you for your interest in our work! I started this research because I felt the limitations of the trajectory prediction field and saw the need to expand into pedestrian dynamics simulation. It would be great if we have a chance to collaborate sometime.

Regarding the GCS preprocessing, I have uploaded all the preprocessed files to preprocessed_datasets.7z in the CrowdES Dataset Zoo. Could you please try testing with this file?

Before using the new preprocessed files, you need to clear the cache created by the dataloader. (Especially for GCS, the simulation dataloader creates a cache because it can use up to 200GB of RAM for 15 hours.) Please delete the cache pickle files in the datasets/preprocessed/gcs/ folder before running.

Crowd-Behavior-Generation/utils/dataloader/simulator_dataloader.py

Lines 357 to 364 in 093e956

	# Load from cache if exists
	cache_path = os.path.join(config.dataset.dataset_preprocessed_path, 'cache', f'{phase}_simulator.pickle')
	if os.path.exists(cache_path):
	with open(cache_path, 'rb') as f:
	cache = pickle.load(f)
	for key, value in cache.items():
	setattr(self, key, value)
	return

[XieShaocong33Ethan]

Thank you very much for your kind reply and for sharing the preprocessed files. I will test them as you suggested.

I really appreciate your work — I believe your approach can be well-suited for simulating daily pedestrian dynamics. I noticed from your model description that you incorporate Markov states, which makes pedestrian behaviour extensible. This could potentially be combined with the pedestrian dynamics model to further reduce collision rates.

Pedestrian simulation is a very important and challenging topic. I am saddened by the tragic crowd crush incident that happened in Seoul, 2022. Predicting crowd movement to prevent such accidents has always been a major goal in this field. However, it is difficult because behaviors differ greatly between low-density and high-density scenarios. In particular, emergent group behaviors often arise as density changes.

I apologize that I am quite busy towards the end of the year, but I will come back to this. In the meantime, may I reach you via the email address listed in your paper?

[InhwanBae]

Thank you for your thoughtful follow-up.

I agree with your points. I introduced Markov states to ensure diversity and consistency in the generative model, but the collision rate remains a significant challenge, especially as explicit collision events are largely absent in the dataset. It's also an interesting point that you mention 2022; I also began developing this research idea that year.

I think it would be a fascinating research topic to re-implement classical concepts—like 'comfortable' or 'emergency' values—within a deep learning framework, while penalizing collisions and modeling the different crowd behaviors that emerge in low- and high-density scenarios. Please feel free to reach me via the email address in my paper whenever you are comfortable. I look forward to hearing from you!