For a while now I’ve been kicking around the idea of a project involving IMU data (X,Y,Z timeseries) for Human Activity Recognition (HAR). While there is a considerable amount of research available online discussing various architectures and approaches, I’ve been unable to find anything regarding pre-trained models. I was wondering if anybody has any experience with this, as the alternative would be to start from scratch with something like the UCI smartphone dataset.
In my research I found Apple’s MLCreate which has an ActivityClassifier available through the turicreate module, and from what I can tell that itself is a pre-trained network, so it seems to be possible. How to get the model parameters out of it though, I’m not sure. I don’t have any experience in the Apple ecosystem but you can export the model in their proprietary format “.model” or “.mlmodel”, so I would assume they are serializing the architecture somehow.