search

raspstephan / nwp-downscale

MIT License
38 stars 9 forks source link

Check log-transform and weighted sampling #45

Closed raspstephan closed 3 years ago

raspstephan commented 3 years ago

Make sure everything works as intended and plot different version to see the impact.

raspstephan commented 3 years ago

After checking, I realized that I made a mistake by computing the weights after the log transform. I now fixed this. In fact, I changed the weights to be computed based on coverage rather than mean precipitation amount. Then I "softened" the weight distribution a little. See #48

Now testing what happens to GAN training with this.

raspstephan commented 3 years ago

GAN training is better with log transform. Still not sure this is the final solution but closing for now.