```custom-WatchMotionIndex
activePost: 3
```

I'm still collecting data for the model, but I thought I should try to do some on-device inference to get a feel for how it performs and if it's fast enough. I used the training data I have collected so far and training a very naive model without much preprocessing. The model is trained on a window of 500 samples (5 seconds at 100Hz), and I run the inference every 100 samples (using the latest 500 samples).

<iframe
  className="video-1-2 mb-8"
  src="https://www.youtube.com/embed/dV49P3sZiJg?si=iq54NC3G1wdatGE3"
  title="YouTube video player"
  allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture"
  allowFullScreen
></iframe>

To get started, I downloaded the data from DynamoDB with a simple Python-script and saved it to disk as JSON files. I then did some very simple preprocessing, just normalizing the data and splitting it into training and validation sets. I'm using CreateML to train the model, it's great for getting started but gives very little control over the training process so I will probably switch to TensorFlow or PyTorch later. This is what it looks like after training for 20 epochs:

![{2400[x]1318}](/journal/241117/watch-motion-first-try-with-on-device-inference-2.png)


First try with on-device inference

Watch Motion

![{2400[x]1600}{priority}](/journal/241116/infinite-mass-at-pustervik-1.jpg)

![{2400[x]1600}](/journal/241116/infinite-mass-at-pustervik-2.jpg)

![{2400[x]1600}](/journal/241116/infinite-mass-at-pustervik-3.jpg)

![{2400[x]1600}](/journal/241116/infinite-mass-at-pustervik-4.jpg)

More photos and prints at [Rockfoto](https://rockfoto.nu/artists/infinite-mass/20241116/30181)

Read the review at [GP.se](https://www.gp.se/kultur/musik/konsertrecension/recension-infinite-mass-pustervik-lordag.07ad24d1-42ff-4fdd-9708-0efec8f8bdd6)


Infinite Mass @ Pustervik

```custom-WatchMotionIndex
activePost: 2
```

Now to the fun part: Getting a lot of training data! I'm trying to collect at least a couple recordings of each exercise every day. My plan is to collect enough data to train a simple model that I can implement in the watch app and do some basic exercise classification on device just to see that it works.

When building a machine learning model to classify exercises, the quality of the dataset is just as important as the algorithm. A balanced dataset ensures that the model learns equally from all exercise types, rather than favoring one over the others. To measure how balanced the dataset is, we can use the coefficient of variation (CV) which simply is the standard deviation divided by the mean. We want the CV to be as low as possible, if the dataset is balanced, the CV will be close to 0.

```custom-WatchMotionMoreTrainingData

```

The data is updated every time I add a new recording to the dataset, so check back later to see how it progresses!
