RBF Gamma

This notebook tries to find the standard deviation, and hence the $\gamma$, for RBF kernels on the RGZ dataset.



In [12]:

    
import h5py, numpy, sklearn.metrics.pairwise, sklearn.cross_validation

TRAINING_H5 = '../training.h5'



In [21]:

    
with h5py.File(TRAINING_H5, 'r') as f:
    features = numpy.hstack([f['astro'], f['cnn_outputs']])
    features, _ = sklearn.cross_validation.train_test_split(features, train_size=0.4, stratify=f['labels'])
    distances = sklearn.metrics.pairwise.pairwise_distances(features, n_jobs=-1)
    sq_distances = (distances**2).reshape((-1,))
    low, med, up = numpy.percentile(sq_distances, [25, 50, 75])



In [22]:

    
1/low, 1/med, 1/up









    Out[22]:





(9.1663069780085295e-05, 2.2006164003589744e-05, 2.5167104710154812e-06)



In [20]:

    
1/low









    Out[20]:





9.3046119711918889e-05



In [ ]: