I am working with a Fourier dataset where the dynamic range of the data is huge, of the order 1e6 to 1e-6. The ResNet architecture I am using posed as a Image reconstruction problem, just cannot optimise the loss function and is in the range of 5e8. It seems like it will take forever to train.
Can I model the data in different way (like normalisation ? ) But am little skeptical on how normalising the Fourier data would impact the MRI set.
I did a grid search for learning rate it came 1000 for initial few epochs, but cost function shot up in trillions. So, are there more methodical way of searching learnt rates sort from cyclical rate ?
If any of you have faced a data , with such a huge dynamic range, how did you make the network learn ?
I personally feel it is quite relevant to the problems people are solving, so posed the issue here.
You can “reverse engineer” your HDR ft image
Creating your own “inverse tone map function” that splits it into three (or more) LowDynamicRange images.
You can use my sample notebook as starting point:
Here I convert an audio signal into three channel spectrogram.
You can create a function that’s converts a ft hdr image (I assume is a single channel) into a three LDR channels image