Train on 25% and 50% - Githubissues

chengsoonong / crowdastro

Cross-identification of radio objects and host galaxies by applying machine learning on crowdsourced training labels.

MIT License

13 stars 1 forks source link

Train on 25% and 50% #224

Open chengsoonong opened 7 years ago

MatthewJA commented 7 years ago

Train a CNN on 25% and 50% of the full RGZ set. Test on RGZ & Norris with Norris labels. See if there's a difference.

MatthewJA commented 7 years ago

For LR, not for a CNN, but here's a plot of accuracy against the training set size (in radio objects). Looks like we need at least 100 radio objects to peak for LR — that's on the same order as the number of objects cross-identified by Norris et al. I suspect the CNN will converge slower.