liquid_time_constant_networks

Why we included this project

This is the official research code behind Liquid Time-Constant Networks, a recurrent architecture that models sequences as continuous-time dynamics rather than discrete steps. That distinction is worth knowing about if you work on forecasting or sensor-based problems, where samples arrive at irregular intervals, noise is common, and you want the temporal structure to stay interpretable. The repo ships ready-to-run training scripts for five time-series benchmarks: hand-gesture segmentation, room occupancy, human activity recognition, traffic volume, and ozone forecasting. Each script takes a model flag, so you can compare LTCs directly against LSTMs, GRUs, and neural ODE baselines on identical data. Researchers and engineers prototyping RNN variants will appreciate the self-contained Python modules and the clear hyperparameter options for reproducing the paper's results. Just note the code targets TensorFlow 1.14, so treat it as a reference implementation and reproducibility companion; the maintained PyTorch port lives in the related ncps project.