Train relaxed models on smoke plume for different alphas. Use rotation equivariance full and partial.

[Nesta-gitU]

Do what the title says, use the small equivariance test dataset.

for part of this issue add a test set to that dataset.

use level 0 and 3 or something

[dgcnz]

done