Oddly enough (although perhaps more common in the HN community) I did this in Tensorflow for a research project with a local aquarium. In the end after normalization, accuracy was about ~78% with ~100k epochs on a sample size of about 400 images per distinct species. Glad to see more ML work in the marine biology/fishing world, which appear to have relatively small annotated datasets compared to other fields.
I liked the article a lot, but the title is a bit confusing. This is a beautiful project about a beautiful algorithm that you implemented. Do you really want to put the name of a library that you used in the title? It takes attention away from the actual problem.
Did you ever open source it? I’m looking for something similar to this!