Closed b-remy closed 1 year ago
I did a longer test using the default flags of the VAE_SD_C.py
script, so most likely not the best you found @JonnyyTorres, so the result are not as good as yours. But it loos like we don't see the convolution pattern anymore
Hi @JonnyyTorres ,
This PR is a attempt to fix the pattern we get in #5 . It adds a new convolution function which takes the PSF in Fourier space as input. I updated the dataset configuration file and the training script
VAE_SD_C.py
to make a few tests.I couldn't perform a full training with a big dataset, but I didn't see any similar pattern in my tests.
You can already see that convolving galsim images with their associated PSF gives better results with the new code:
first convolution code
new convolution code
I commented the lines to register galaxies noise standard deviation to make faster experiments. I would be interested to see a full training with this configuration and make to complete dataset with the correct std in a second step.