Federated-Wearables is a demo for cross-device federated learning over wearables. This is the code of the smartwatch which offloads the collected data to the paired phone via Bluetooth for training in FL, and gets the updated model to run interface in real time.
In this demo, we conduct an activity recognition task. To coordinate the FL and alleviate the calculation burden of wearables, we developed a companion app in Application directory. Notably, the app enables users manually correct the labels to generate more precise training samples upon witnessing wrong interface given by the model.
Feel free to use this demo for your own task or application:)
The inference happens directly on smartwatchs.
All is in Trainer/Wearable/src/main/java/fl/wearable/autosport/MainActivity.java
In the folder Trainer/Wearable/src/main/java/fl/wearable/autosport/sensors. Here we implement sensor codes for those available in watches, such as gyroscope, heart rate, geolocation, etc.
We manually implement the inference of a Multi-Layer Perception classifier in Trainer/Wearable/src/main/java/fl/wearable/autosport/model/Classifier.java.
The training requires the coordinations of a parameter server, the android app on the phones.
In the federated training, the model checkpoint is saved in Trainer/Application/src/main/assets, which is loaded by the smartwatch in the beginning of inference. The details of the parameter server is in Trainer/Application/src/main/java/fl/wearable/autosport/.
KotlinSyft and TrackerPublic.