Fully Convolutional Network Bootstrapped by Word Encoding and Embedding for Activity Recognition in Smart Homes

Code

The code is divided as follows:

• The main.py python file contains the necessary code to run an experiement.
• The classifiers folder contains deep neural network models tested in our paper.
• The pre_proccessing folder contains scripts to preprocess datasets.

Data

The data used in this project comes from Center for Advanced Studies in Adaptive Systems CASAS :

• The Aruba

• The Milan

• datasets contain a copy of the original datasets and an exemple allready pre-processed with a sliding windows of size 100 and 25.

Setup

To quickly run a training, extract the pre-processed datasets in datasets folder and see the section "Train" below

Pre-processing

First step is to read raw dataset pre-process it with pre_process.py.

• STEP 1: Load dataset
• STEP 2: prepare dataset
• STEP 3: segment dataset in sequence of activity
• STEP 4: transform sequences of activity in sentences
• STEP 5: sentences indexization
• STEP 6: split indexed sentences in sliding windows
• STEP 7: pad sliding windows
• STEP 8: save sliding windows and labels

python pre_process.py --i ../datasets/original_datasets/milan/data --o MILAN_activity_sequences_windowed --w 25  

Options

• --i input file dataset
• --o output file
• --w windows size

In order to train models on equal bases, uses subsets_generator.py. This script generate train, validation and test subsets.

python subsets_generator.py --i MILAN_activity_sequences_windowed_25_padded --p ../datasets/MILAN    

Options

• --i input file dataset
• --p path where subsets will be save

Train

To run models on one dataset you should issue the following command:

python main.py --i MILAN_activity_sequences_windowed_25_padded --d datasets/pre_processed_datasets/MILAN --c batch_1024_epoch_400_early_stop_mask_0.0 --m FCN_Embedded FCN 

Options

• --i input file dataset
• --d path to the dataset
• --c comments ex: choosen hyperparameters, batch size, ...
• --m models one ore more

Results

The following table contains the averaged balanced accuracy and average weighted F1-score over 3 runs of each implemented models on Aruba and Milan datasets from CASAS depends on the sliding windows size used.

			ARUBA					MILAN
Metrics	Models	100	75	50	25		100	75	50	25
Weighted avg F1 Score (%)	LSTM	96.67	94.67	90.67	85.00		84.00	85.67	75.33	64.00
	FCN	99.00	98.00	97.67	92.33		77.33	93.67	88.33	83.67
	LSTM + Embedding	100.00	99.67	98.00	90.00		98.00	97.00	93.00	73.67
	FCN + Embedding	100.00	100.00	100.00	99.00		99.00	98.00	97.00	94.33
Balanced Accuracy (%)	LSTM	81.45	76.09	71.05	83.30		62.15	64.95	55.70	43.29
	FCN	88.85	87.41	87.08	80.32		42.24	76.41	71.82	71.34
	LSTM + Embedding	94.55	93.61	90.20	74.81		88.52	86.77	82.05	59.35
	FCN + Embedding	95.37	95.07	94.89	92.44		84.23	86.64	87.83	90.86

Reference

If you re-use this work, please cite:

@article{dbouchabou2020fcnsmarhomehar,
  Title                    = {Fully Convolutional Network Bootstrapped by Word Encoding and Embedding for Activity Recognition in Smart Homes},
  Author                   = {Damien, Sao Mai, Christophe, Benoit, Ioannis},
  journal                  = {},
  Year                     = {2020},
  volume                   = {},
  number                   = {},
  pages                    = {},
}