Train a CNN on FIRST

[MatthewJA]

FIRST is way too big to store all of in an HDF5 file, so we'll have to do feature extraction on the fly. This probably needs a deeper CNN model than the one we're using, and/or PCA, just to get the sizes down a bit.

I propose the following: Sample 200 FIRST objects with crowd labels, get the radio patches associated with the nearby galaxies, and train a 3-convolutional-layer CNN on just this. Then PCA the result and use the resulting pipeline to extract features while importing FIRST.

If this is a decent idea, I can throw together a script to run on a GPU and get a dataset together.

[chengsoonong]

Please don't do anything about training on FIRST until all your other results are done. We should have a chat about how to best proceed when we next meet.

[MatthewJA]

Alright, sounds good to me!

[MatthewJA]

I'll put all the related issues into a label so that we can easily find them later.