For annotation, check out Prodigy [0]. Generally speaking, as classification systems themselves are pretty dumb there isn't really a way to know what architect

For annotation, check out Prodigy [0]. Generally speaking, as classification systems themselves are pretty dumb there isn't really a way to know what architecture will work best for your task, other than trial and error. Of course you can optimize parameters in a less chaotic way (grid-search or AutoML). In my experience it mostly boils down to data. Try augmentation methods, acquiring more data or transfer learning with varying degrees of layer relearning. [0]: https://prodi.gy/ https://prodi.gy/