Tal Shiri

~Hey - I am experiencing the same issue (nans after about 200 iteration steps). `perspective_correct=False` doesn't seem to help though :(~ EDIT: I didn't notice they were multiple RasterizationSettings instances....

> As a workaround for 0.50, you could try setting `"remote.SSH.useLocalServer": false`, and let me know whether that helps. Can confirm this works

> Sorry for the trouble everyone. I can't figure out what fish is doing here to break this. I just spawn an ssh process and pipe a script into stdin,...

> Sorry for the trouble everyone. I can't figure out what fish is doing here to break this. I just spawn an ssh process and pipe a script into stdin,...

@roblourens ah I see. You could run a probe for the platform in a preflight call. This is stable, because stdin would get closed as soon as `uname -rsv` gets...

> is this related? [fish-shell/fish-shell#1439](https://github.com/fish-shell/fish-shell/issues/1439) ha maybe. Good find!

> Many things would be simpler if I could run multiple ssh commands, but I have managed to keep it to a single ssh connection at least per window, and...

> @jcpeterson My guess as to why the encoder works better as-is, is that it has the full latent space to search through, so while the mapping network is designed...

To illustrate the last point, here's a simple ablation test (this is all done in `W{1}`): target_image:  mask:  Target = target_image * mask (so only the white pixels...

> @tals I've also been [playing around with encoders and training a ResNet to encode](https://github.com/pbaylies/stylegan-encoder); I had some similar ideas as far as using the mapping network to generate more...