UniMTS: Unified Pre-training for Motion Time Series

Authors: Xiyuan Zhang, Diyan Teng, Ranak Roy Chowdhury, Shuheng Li, Dezhi Hong, Rajesh Gupta, Jingbo Shang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments We evaluate on the most extensive motion time series classification benchmark to date, comprising 18 real-world datasets that cover diverse activities.
Researcher Affiliation Collaboration Xiyuan Zhang UC San Diego xiyuanzh@ucsd.edu Diyan Teng Qualcomm diyateng@qti.qualcomm.com Ranak Roy Chowdhury UC San Diego rrchowdh@ucsd.edu Shuheng Li UC San Diego shl060@ucsd.edu Dezhi Hong Amazon hondezhi@amazon.com Rajesh K. Gupta UC San Diego rgupta@ucsd.edu Jingbo Shang UC San Diego jshang@ucsd.edu
Pseudocode No Not found. The paper describes processes and frameworks but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes 1Code is available on Github: https://github.com/xiyuanzh/Uni MTS. Model is available on Hugging Face: https://huggingface.co/xiyuanz/Uni MTS.
Open Datasets Yes We simulate motion time series from existing motion skeleton dataset Human ML3D [19], which contain both motion skeleton data and corresponding text descriptions as detailed in Section A.1 in Appendix.
Dataset Splits No Not found. The paper discusses train, few-shot, and zero-shot settings but does not explicitly specify a validation dataset split or how it was used.
Hardware Specification Yes We pre-train Uni MTS using Adam optimizer [25] with a learning rate of 0.0001 on a single NVIDIA A100 GPU.
Software Dependencies Yes We prompt GPT-3.5 ( gpt-3.5-turbo ) to generate k = 3 paraphrases.
Experiment Setup Yes We pre-train Uni MTS using Adam optimizer [25] with a learning rate of 0.0001 on a single NVIDIA A100 GPU. The pre-training process consumes approximately 13 GB of memory given a batch size of 64. For text augmentation, we prompt GPT-3.5 ( gpt-3.5-turbo ) to generate k = 3 paraphrases. During each iteration, we randomly generate the mask M by selecting 1 to 5 joints and mask the remaining joints as zeros. We adopt learnable temperature parameter γ initialized from CLIP.