Human Activity Recognition with improved LSTM models

Performance comparison based on CNNLSTM, Convolutional LSTM, Stacked LSTM, Residual LSTM on Human Acitivity Dataset UCI_HAR_Dataset

Environment Requirements

Miniconda

Anaconda

• python 3.6.7
• Tensorflow 1.12.0
• keras 2.2.4

For StackedLSTM and ResLSTM, the experiments must be executed on one GPU with CUDA support.

Explaination for every function file

filename	Functionality
allModels.py	The definition for 4 different models
all_utils.py	The different supplementary infrastructure functions we used
runExperiment.py	The execution of the experiments

Running the experiments

To run the program, activate proper environments first:

source activate keras_env

Flags available to modify our experiment setting

arguments	setting
repeats	The number we repeat our experiments
grid	Weather execute grid search or not, set True or False
arch	The choice of model architecture, chosen in ['conv_lstm', 'cnn_lstm', 'res_lstm', 'stacked_lstm']

start training and test one model with a command like this:

python runExperiment.py --grid True --arch conv_lstm --repeat 5