Machine learning model for classifying transportation mode

HansBambel commented 4 years ago

This PR adds:

[x] Dataset class for transportation modes
[x] Function to create train+valid and test set folders from data
[x] Dataset are normalized using mean and std of training set
[x] Neural Network with Convolutions ("A Convolutional Neural Network for Transportation Mode Detection Based on Smartphone Platform")
[x] Using all sensors or (like in the above paper) magnitude of accelerometer with removed gravity
[x] Working randomforest classification
[x] Shrunk _accmagnitude dataset creation from ~11-17s per file to merely 0.3s per file (thanks to numba)
[x] Apply numba to normal dataset creation
[ ] Smoothing of signals (postpone to later)

Added need for numba-package. This reduces creation of magnitude calculation to only 10% of pure python implementation.

HansBambel commented 4 years ago

ROC score of Randomforest (with 100 trees) is 0.72 on the (non-corrected) blackforest dataset

HansBambel commented 4 years ago

Smoothing requires scipy.signal.savgol_filter to be able to converted to numba function... maybe later...

dimstudio / SharpFlow