Open wty-ustc opened 3 years ago
So the (k, v) correspondence comes from the "copy-paste" action.
See the function rewrite.ganrewrite.ProgressiveGanRewriter.paste_from_selection
. The output variable goal_in
corresponds to k and the variable goal-out
is v. Note that goal_in
is the context_model
(layer "L-1") representation of the target context, like a building spire, and the goal_out
is the target_model
(layer "L") representation of the replacement pattern, like a tree that we want to put in place instead of building spires. Note that we can get both representations through just inference in the GAN: no GAN inversion is needed.
In you paper,
I am confused how to get v*i?
Through gan inversion optimization to find the vi correspond to x_i?
I can't find in your code, so sorry to bother you.