search

weecology / DeepTreeAttention

Implementation of Hang et al. 2020 "Hyperspectral Image Classification with Attention Aided CNNs" for tree species prediction
MIT License
119 stars 38 forks source link

train a representation space before classifier #84

Closed bw4sz closed 2 years ago

bw4sz commented 2 years ago

Use all data from all species to train a representation space to classify within. There are thousands of data that get removed during train/test filters.

bw4sz commented 2 years ago

Can we use the same representation space as the autoencoder for filtering samples? Remove test dataset first? This somewhat conflicts with novel species testing. I like @MarconiS known unknown idea, i'm not sure if there is a way of including the novel class in a supervised setting, since by definition, they are outside of the classifier. I am worried about leak from test dataset if this is too closely tied to the autoencoder.

bw4sz commented 2 years ago

I'm not sure where 'metric' learning begins and 'representation learning' ends. Need to ask alina.