My team is anxious to start using the FTC-ML system. Its been difficult to understand where to go for the latest information on FTC-ML so we continue to monitor the following sites for any news and information about using this ML toolchain.
Github Wiki - FtcRobotController/wiki/Using-TensorFlow-in-Freight-Frenzy
Is there a definitive channel we subscribe to or monitor to know when this system is available for use?
Per the manual our team is registered and roles are assigned in ftc-scoring system but all we see is the Getting Started/ML Resources links (see attachment)
Please help us fill in our gaps in understanding how to get started.
At minimum it would be great to know if this will be available in the next week. The team is concerned that we’re running out of time to take advantage of ML for upcoming competitions.
I apologize for the lack of communication in this respect - our intent was to make the FTC-ML tool available last week, but TensorFlow released a new version on or about November 16 that exposed a problem within the FTC-ML tool that caused widespread issues within this tool. We did not notify the community about this problem because the Beta Test was officially “over” even though we did not “shut down” the system; the issue related to the use of our accelerators (GPU or TPU accelerators). Models can be created but instead of processing about 1 step per second (which is about the average) it is taking upwards of 4 seconds to process a single step. This kind of issue not only wreaks havoc on our accounting within the system but it also costs us more in system resources to create models. So we’ve been hard at work trying to fix this so that we can open it up to the broader community while not bankrupting us in the process. We’re making great progress thanks to the efforts of engineers at Google giving us a LOT of help (and of course Liz who has been our Rock Star in this entire process).
It is our hope to open the tool within the next week, the locations you’re watching are all great locations - we’ll be announcing the opening of the tool via the FTC Blog, this FTC-ML forum, and via team e-mail blasts.
If you want to get started, there’s two ways to get started:
(1) If you want to perform the full ML process, you can begin reading the FTC-ML User Manual and begin taking video of your object.
(2) If you want to shortcut the ML process, download a Zoo Model (as described in Section 7.1.2), print out an image from the COCO training data and tape it to your custom object (or even use a red solo cup), use Section 6 to modify the custom model template to use the Zoo Model and add the proper label, and begin tracking your custom object using the image as a reference.
