Human activity recognition with pytorch-lightning

This is a Pytorch implementation of the basic CNN approach proposed in there: https://machinelearningmastery.com/cnn-models-for-human-activity-recognition-time-series-classification/

How to run

First, install dependencies

# clone project   
git clone https://github.com/YourGithubName/Your-project-name   

# Create and activate virtual env (OPTIONAL)
cd human-activity-recognition
python3 -m venv har
source har/bin/activate

# Install dependencies
pip install -r requirements.txt

Next, download the dataset and unzip it in a data/. Your folder should look like that:

.
├── README.md
├── data
│   └── UCI HAR Dataset
│       ├── README.txt
│       ├── activity_labels.txt
│       ├── features.txt
│       ├── features_info.txt
│       ├── test
│       └── train
├── requirements.txt
├── setup.py
└── src

Next, navigate to the baseline code and run it.

# module folder
cd src/cnn1d/baseline

# run module (example: mnist as your main contribution)   
python trainer.py    

With the proposed setting, after 10 epochs we obtain the following confusion matrix and a validation accuray of 90.7%:

[[462  16  18   0   0   0]
 [  6 439  26   0   0   0]
 [  2   5 413   0   0   0]
 [  0   9   0 375 107   0]
 [  1   2   0  70 459   0]
 [  0  14   0   0   0 523]]

Exploration

We have tried to replace the CNN with an RNN which should be more adapted to temporal data. the files can be found in src/rnn:

# module folder
cd src/rnn/baseline

# run module (example: mnist as your main contribution)   
python trainer.py    

The results are not as good as the baseline CNN, we achieve a validation accuracy of 80%.