Policy information about availability of computer code.
This document lists the software used in this study.
Provide a description of all commercial, open source and custom code used to collect the data in this study, specifying the version used OR state that no software was used.
No software was used for data collection. This study uses pre-existing video datasets (e.g., BabyView); no new data was collected.
Provide a description of all commercial, open source and custom code used to analyse the data in this study, specifying the version used OR state that no software was used.
All analyses were performed with custom code written in Python 3.10, available under the MIT license at https://github.com/awwkl/ZWM, with pretrained model weights at https://huggingface.co/awwkl/models.
Model training and inference used PyTorch 2.8.0, torchvision 0.23.0, and Triton 3.4.0, with CUDA 13.0 / cuDNN 9.19. Numerical and scientific computing: NumPy 2.2.6, SciPy 1.15.3, pandas 2.3.3, einops 0.8.2, einx 0.4.3. Image and video I/O: scikit-image 0.25.2, Pillow 11.3.0, opencv-python 4.13.0.92, decord 0.6.0, moviepy 2.2.1, imageio 2.37.3, h5py 3.16.0, tifffile 2025.5.10. Machine learning utilities: scikit-learn 1.7.2, vector-quantize-pytorch 1.28.2. Visualization: matplotlib 3.10.8. Experiment tracking with Weights & Biases (wandb) 0.26.0. Pretrained model distribution via huggingface_hub 1.11.0 and google-cloud-storage 3.10.1. Interactive demos built with Gradio 6.12.0.
A complete pinned dependency list is provided in requirements.txt for full reproducibility.