Liquid time-constant Networks (LTCs)

[Update] A Pytorch version is added in our sister repository: https://github.com/mlech26l/keras-ncp

This is the official repository for LTC networks described in paper: https://arxiv.org/abs/2006.04439 This repository alows you to train continuous-time models with backpropagation through-time (BPTT). Available Continuous-time models are:

Requisites

All models were implemented tested with TensorFlow 1.14.0 and python3 on Ubuntu 16.04 and 18.04 machines. All following steps assume that they are executed under these conditions.

Preparation

First we have to download all datasets by running

source download_datasets.sh

This script creates a folder data, where all downloaded datasets are stored.

Training and evaluating the models

There is exactly one python module per dataset:

• Hand gesture segmentation: gesture.py
• Room occupancy detection: occupancy.py
• Human activity recognition: har.py
• Traffic volume prediction: traffic.py
• Ozone level forecasting: ozone.py

Each script accepts the following four agruments:

• --model: lstm | ctrnn | ltc | ltc_rk | ltc_ex
• --epochs: number of training epochs (default 200)
• --size: number of hidden RNN units (default 32)
• --log: interval of how often to evaluate validation metric (default 1)

Each script trains the specified model for the given number of epochs and evalutates the validation performance after every log steps. At the end of training, the best performing checkpoint is restored and the model is evaluated on the test set. All results are stored in the results folder by appending the result to CSV-file.

For example, we can train and evaluate the CT-RNN by executing

python3 har.py --model ctrnn

After the script is finished there should be a file results/har/ctrnn_32.csv created, containing the following columns:

• best epoch: Epoch number that achieved the best validation metric
• train loss: Training loss achieved at the best epoch
• train accuracy: Training metric achieved at the best epoch
• valid loss: Validation loss achieved at the best epoch
• valid accuracy: Best validation metric achieved during training
• test loss: Loss on the test set
• test accuracy: Metric on the test set

Hyperparameters

Trajectory Length Analysis

Run the main.m file to get trajectory length results for the desired setting tuneable in the code.