Open SergeKrier opened 7 years ago
in initial implementation, we totally missed the fact that original image size : 1) varies -> need to image size as feature 2) size is usually smaller than 224x224 (for vgg/resnet) , thus need upscaling -> use cv2 linear or cubic interpolation
Currently we use:
df['lin_mass'] = np.power(10, df.logMstar)
df['lin_err'] = df.lin_mass * np.log(10) * df.err_logMstar
err_lin_to_L = df.lin_err.values[:N]/(df.LtoFlux[:N]*Xflux)
Current gold solution is using linear interpolation, I haven't tried cubic yet.
As per Marteen's email, we should transform the mass to mass to luminosity ratio, and luminosity = 4 pi d^2 * flux. Then we should remove distance from the features.