-
Week six: training the CNN for the first time, and the LSTM that should have beaten it
I trained both models myself this week. The one that was supposed to be better came out worse, and I'd rather show that than the flattering curve.
- Week six: hardware at Makerspace, and a new teammate The cube is nearly designed, its parts are on a bench at Makerspace, and there is one more of us.
-
Week six: both models trained, and one result I can't explain yet
I trained the CNN for the first time and the LSTM on top of it. The model that should have won didn't, and I'd rather show that than the flattering curve.
- Week six: moving the dashboard off Streamlit A decision, not a writeup: the carer dashboard is moving off Streamlit, because an alert that arrives late is worse than no alert.
-
The other branch: what patient detection does when the face isn't enough
Pulse has a second branch for when the patient is out of bed. It reads the body instead of the face. Here is what the demo shows today, and what it doesn't yet.
-
We checked whether pain has a shape before we built a model that assumes it does
Before training anything, I plotted 48,398 frames of the dataset. One finding validated our architecture. One we cannot engineer our way past.
-
Writing the contract between our CNN and our LSTM, before either exists
Two people are about to train two models against two different ideas of what a feature vector is. Here is the contract we wrote to catch that in advance.
-
Our dataset is a physio clinic. Our room is a bedroom at 3am.
What UNBC-McMaster actually contains, and the honest distance between that footage and the room we are aiming at.
-
Skeletons before data: writing the loop ahead of the models
We wrote the whole inference loop with stub models returning random numbers. It caught four interface bugs before the weights existed.
-
Why the device has a face
A camera pointed at a bed is a surveillance object until you give it a reason not to be. The face is that reason.
-
Three models, not one: our plan for splitting the pipeline
One end-to-end network would be simpler to draw and much harder to debug. We are choosing debuggable, before we've trained anything.
- Week one: we argued our way from falls to pain Falls detection was the obvious project. Two hours of arguing got us somewhere less crowded and more useful.