Bug description

In SLEAP v1.5.1, training a model with the original label .slp file is very slow (0.5 it/s). Using the exact same config, but training on an exported .pkg.slp file, the speed is normal (~4 it/s). Interestingly, when running the same training from sleap v1.4.1, I do not run into the issue of slow training. I assume it must be something related to having a high number of videos in my label file and data loading.

I run

uv run sleap-nn-train --config-name centroid.yaml  --config-dir ~/.../sleap_files/training_251020_1453

I attached my config file. To compare, I just change the path from the .slp to .pkg.slp (see comments).
centroid.yaml

System

Ubuntu 20.04.6 LTS
RTX 3060 12GB, Driver 560.35.05, Cuda: 12.6.

Installation method:

mkdir sleap
cd sleap
uv venv --python 3.12
uv init
uv pip install "sleap[nn]" --index https://download.pytorch.org/whl/cu126 --index https://pypi.org/simple```