something wrong with navsim mini dataset

[Bronyaer]

Environment

• Dataset: OpenScene mini
• Split: navmini
• GPU: RTX 4090 (24GB)

Issues found

1. scene_filter_path not available for mini split

In train_caching.py and run_dataset_caching.py, the code assumes
cfg.train_test_split.scene_filter_path always exists, but the mini
and navmini splits use an inline scene_filter config instead of a
msgpack file path.

Fix:

if hasattr(cfg.train_test_split, 'scene_filter_path'):
    scene_filter_cfg = load_msgpack(cfg.train_test_split.scene_filter_path)
    scene_filter_obj = instantiate(scene_filter_cfg)
else:
    scene_filter_obj = instantiate(cfg.train_test_split.scene_filter)

2. train_val_test_log_split_path incompatible with mini dataset

The default default_train_val_test_log_split.msgpack contains logs
from the full navtrain/navtest splits. None of the navmini logs appear
in val_logs, so the validation dataset is always empty during training.
This causes val/score_epoch to never be logged and ModelCheckpoint
to not work correctly.

Could you provide a navmini-compatible split file, or update the
training script to handle this case?

3. Hardcoded VoVNet checkpoint path

In gaussianfusion_config.py:

vov_ckpt = "/home/liush/GaussianFusion/pretrained/dd3d_det_final.pth"

This path is hardcoded to the author's local machine. Please make
this configurable (e.g., via the yaml config or an environment variable).

4. Training scripts use navtrain by default

run_gaussianfusion_training.sh uses TRAIN_TEST_SPLIT=navtrain,
but users with only the mini dataset cannot use this.
It would be helpful to document how to run with the mini dataset.

Question

Could you provide documentation or a guide for reproducing with the
mini dataset? Thank you!