Evaluation Module

[YHRen]

To compare the output of GAN and "real" simulation results.

[YHRen]

see cosmoflow
'predict_Cosmo.py`

[YHRen]

It seems the loss is the pixel-wise "mean_squared_error"(L2) between generated image and ground truth.
This is common during training, but I'm expecting something like mutual info on the histogram of the spatial density of the universe.
Because different (random) initial conditions would lead to very different pixel-wise outcomes.

The L2 makes sense only if the starting point of GAN is an initial condition from MUSIC, rather than a set of parameters.
Namely, the CosmoGAN is only the surrogate model for PyCola, but not MUSIC+PyCola.