This part of the repository contains python scripts and notebooks to run the hsmm4acc code on the Millenium Cohort data set. It finds states in the accelerometer data with the unsupervised Hidden Semi-Markov Models.
- Python 2.7
- For the notebooks: iPython
- Python package hsmm4acc (see GitHub page for installation instructions)
- packages from requirements.txt
The file config.yml defines the directory in which data is stored, as well as all parameters for the HSMM models. The scripts will create subdirectories under the data directory, assuming the following directory structure:
<data_path>
├── accelerometer_5second
│ ├── <output files of R code>
├── merged
│ ├── <output files of step 0-Preprocess>
├── results
│ ├── mod_<n>st_<b>b_<r>r_<t>t_<colnames>
│ │ ├── model.pkl
│ │ ├── config.yml
│ │ ├── datawithstates
│ │ │ ├── <output files of step 1-HSMM>
│ │ └── images
│ │ │ ├── <output files of notebooks>
├── subsets
│ ├── <output files of step 0-Preprocess>
0_prepare_data.py: This script does the following things:
- joins wearcodes with the TUD (Time Use Diary) file
- removes accelerometer data where the data is invalid
- mirrors accelerometer data that seems to be of devices worn upside-down
- joins accelerometer data with TUD files.
1_HSMM.py This script trains the HSMM on all data in the merged folder, it ouputs the data with states in the datawithstates folder.
5_apply_model.py This script is to apply the model to all data in the merged folder.
The notebooks are for analysis of the outcomes of the HSMM model.
1_HSMM.ipynb has the same functionality as the script, but shows a few plots.
2_AnalyseResults.ipynb shows plots and tables that describe the distribution of the variables and states, as well as a comparison to the reported activities from the diary.
3_AnalyzeModel.ipynb shows the distributions learned by the model
4_3dVisualization.ipynb shows a 3D visualization of the angles and states.