2024-04-28
As I discussed here, I abandoned my exploration of the Conformer, at least for now, and went back to version 2 of my Manta Ray Re-Identification project. I am trying to refactor to a cleaner environment and that seems to argue for replacing Detectron2 with another object-detection model (I first find the manta ray(s) in a photo and then do identification on the cropped image). I'm looking at EfficientDet, with a cloud-first focus.
Otherwise, I'm working on my first re-read of "Dune Messiah" since I was a kid. Oh my god, is it terrible so far, two-thirds of the way in. This novel could have been an email: "Paul broods about his future."
2024-04-14
- The Sharkbots had a fantastic showing at the Hawaii regional, ranking 11th out of 42 teams and making it deep into the championship rounds. This is a really tremendous achievement for such a small, poorly funded school and I am working to try to maintain some of that energy with further projects for the rest of the school year.
- I did some initial structural work for the BirdCLEF challenge, but my initial architecture (a Conformer, which is a combination of Convolutional and Transformer architectures) takes more than 24GB of RAM for training, so I either have to cut it down considerably or pursue multi-GPU training in the cloud. We'll see.
- One of our rainy seasons has begun. We often get quite gloomy from around now until the very start of Summer. I bring this uo because "what I'm doing" involves lots of house projects I've left moribund: new WiFi, putting a smoke detector in the garage, and so forth. Working on Boaty is difficult on rainy days.
2024-04-04
- First Robotics season coming to an end, with the Sharkbots off to Oahu for this weekend's competition.
- Going to take a run at Birdclef 2024
- Annual maintenance on my 17' Boston Whaler, now that whale season is over