Tips generating custom classifiers

[austinczeller]

I am attempting to make a number of custom classifiers for small mammal species in Northern Canada. I was wondering if there are any general tips or rules of thumb when selecting training data.

The vocalizations I am using are short (collared pika vocalizations: >1 second). Should I crop around the vocalization or should I keep most/all of my training data at ~3 seconds. Also, should I include all levels of quality in my training dataset? For example, I have a bunch of vocalizations which are quieter or occur simultaneously with other species' vocalizations.

My attempts to create a custom classifier so far have been unsuccessful, AUPRC across all epochs equals 1. These classifiers result in a false positive for every three seconds of audio.