With the completion of the NASA Phase I SBIR I thought it would be useful to tag the versions of Myriad, Myriad Desktop, and Myriad Trainer that were delivered at the end of the project. They’ve all been tagged as version “1.0-SNAPSHOT” and here are the links:
- Myriad https://gitlab.com/ccoughlin/datareader/tags/1.0-SNAPSHOT
- Myriad Desktop https://gitlab.com/ccoughlin/MyriadDesktop/tags/1.0-SNAPSHOT
- Myriad Trainer https://gitlab.com/ccoughlin/MyriadTrainer/tags/1.0-SNAPSHOT
Since we’re likely to break backwards compatibility in the next version, for your convenience here’s the trained model bundle we put together if you want to take a test drive. What’s in the bundle? Quoting from the original docs:
Inside is a model that has been trained to find indications of structural damage in sensor data and some sample input files to get you started. Download the bundle and extract the model and the sample files in a convenient location, then use them as you proceed through the documentation.
Basically once you’ve built the 1.0-SNAPSHOT-tagged source code, this bundle gives you a starting point for training a machine learning model to detect damage in ultrasonic data and using it in a distributed damage detection system.
So what’s coming in the next version of Myriad? First up is a major redo of the serialization. The initial approach to serialization was functional but not easy to use either as a developer or as an end user. This was fine in the Phase I where you’re under time constraints to show feasibility early on, but there’s definitely room for improvement. Among other things I’m hoping to implement is a “black box” of sorts that bundles the machine learning model and its pre-processing functions in a single package so that an end user doesn’t need to keep track of how to massage raw data for the model, just feed it in and get the results. This should make it easier to share and distribute machine learning models.