HAR

Human Activity Recognition using smartphone data.
Dataset: UCI HAR

Algorithm	Precision	Recall	F1-Score
LSTM	0.89195 (+/-0.00917)	0.88775 (+/-0.00913)	0.88788 (+/-0.00881)
CNN-LSTM	0.89570 (+/-0.00866)	0.89121 (+/-0.01087)	0.89127 (+/-0.01103)
ConvLSTM	0.90512 (+/-0.00701)	0.90037 (+/-0.00921)	0.90071 (+/-0.00907)

Benchmark results

python code/run_experiments.py -h

usage: Run models on the UCI HAR dataset. [-h] [--dataset DATASET]
                                          [--repeats REPEATS]
                                          [--models MODELS [MODELS ...]]

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     Root path to UCI HAR dataset
  --repeats REPEATS     No. of repeats for each model
  --models MODELS [MODELS ...]
                        List of models to evaluate on. Valid models are:
                        [lstm, cnn_lstm, conv_lstm, simple_cnn, wavenet_cnn]

Output

Get Precision, Recall and F1 score for the models across 'repeats' runs.
Output will be generated in the folder 'results' and models will be saved in 'models'.

Also, tensorboard compatible training logs are generated for each run under the folder 'logs' and subfolder 'model name'.

LSTM-CRF Model

jupyter notebook code/notebooks/data_loader.ipynb
jupyter notebook code/notebooks/lstm_crf.ipynb

Baseline results

python code/baseline.py

Data Analysis

jupyter notebook code/notebooks/EDA.ipynb

Android App

Project build folder code/HARApp