Third Task - Konnsy/REAML2022-hackathon GitHub Wiki
This task is a specialization of the Second Task. Try to modify your solution to meet updated requirements.
When targeting a mobile usage of your system, you will encounter constraints that will force you to save ressource consumption. In this case, you depend on your system to run on an Odroid N2+. It is an embedded system that is small compared to a regular PC or laptop computer and consumes less energy. It could thus be built right into the sensor and run for a limited time on a battery.
Memory
You have to make sure that your model does not use more memory than available on the device. there are in total 4GB of RAM available for the operating system (already installed) and all parts of your detection system. Effectively, this means that only about 2.6GiB are available for your program. Keep in mind that a lower memory consumption and fewer memory operations (loading, writing, copying data) often result in a lower energy consumption.
Computation Power
The device has only limited computing power and no CUDA-capable graphics card. Try to find an efficient way to use its resources.
Provided Code
Use your code from Task 2 as a starting point. Make sure that test analysis at the end of the code works so that you can produce results for your submission when we give you previously unknown test data sets.
Ressource Measurements
In order to measure the demands of your detection system, have a look at the examples in the code snippets to measure the used time and memory of an execution. Instead of trying your models out on an Odroid directly, you can check beforehand if it is plausible that it will run on it. This can save some time and can help you in the development process.
Testing on site
You can test your code on some of these devices on site. Use them to check if your approach is able to be executed on them. As well as, evaluating assumptions about the speed of operations and of your overall system in the development process. If you want to adapt the system to your need, please ask before doing so.
Scoring
We will measure execution times on a test set directly on the used device. Make sure that your model has an accuracy of above 50% for the validation sets given in task 2.