Set environment variable HAR_PIPELINE_PATH to the root of your copy of https://github.com/davors/HAR-pipline.
You can set up the Pipenv environment to run mobilenet_uci-har.py with the provided Pipfile.
Run the model with the following command. Select whether you want to train the network. Pass a path to the file where you want the PyTorch checkpoint to be created after training. This value is also used when evaluating an trained model.
python mobilenet_uci-har.py train|test path_to_pytorch_checkpoint.pth
The neural networks descriptions found in python scripts were taken from the ApproxHPVM. Direct link to the directory where you can find them.
Data used was from UCI-HAR dataset. Data from Inertial-Signals were used. This data is filtered using median filter. Accelerometer data is split into 2 components: gravitation and body accelerations, using a low/high pass filter with frequency 0.3Hz. Segment vectors are 128 samples long, sampling rate was 50Hz.
Each channel (body_acc, body_gyro, total_acc) is put into a 32x32 single channel image like shown below, resulting in 3x32x32 signal images.
[
8 rows for d in [x, y, z, y, x, z, x, y]:
channel.d[0..32]
8 rows for d in [x, y, z, y, x, z, x, y]:
channel.d[32..64]
8 rows for d in [x, y, z, y, x, z, x, y]:
channel.d[64..96]
8 rows for d in [x, y, z, y, x, z, x, y]:
channel.d[96..128]
]
| epochs | learning rate | accuracy | file |
|---|---|---|---|
| 2 | 0.10 | 74% | mobilenet_uci-har_0.74.pth |
| 5 | 0.40 | 87% | mobilenet_uci-har_0.87.pth |
| 11 | 0.80 | 90% | mobilenet_uci-har_0.90.pth |
