Code project for the kaggle whales competition on kaggle https://www.kaggle.com/c/noaa-right-whale-recognition
The codebase used is by Sander Dieleman for his ndsb winning solution.
Update to lasagne v0.2Dev - DONE
Make a data converter from kaggle whales to plankton style setup - DONE
Rewrite the data loader to not relying on whole dataset in memory - DONE (though as I only use 512x512 it can actually fit in memory, why I use the branch testingreg now)
Train a neural network - Got a top 20% submission
Further I have a bunch of "lead up" projects, that I havn't had time to test out, but could be cool.
Scale SPN(code works, just needs to write config): https://github.com/alrojo/ZoomSPN
SPN B-Trees(code works, just needs to write config): https://github.com/alrojo/recurrent-spatial-transformer-code
Deep residual networks(code works, just needs to write config) https://github.com/alrojo/lasagne_residual_network
Should test out ZCA-whitening of local pixel chunks, like 8x8 or 16x16.