Jonny Taylor

I had this issue. It was because the linear_assignment import from sklearn is depracated and I had replaced it with the scipy equivalent, which is not quite the same. I...

@wesselb that script gives me OK for both gnp and agnp, so all good there I've just been testing with this script below. Very variable results, with convgnp the performance...

@wesselb here's a messy chatgpt script to make 2 identical relu networks, just using pytorch and tensorflow. It seems to train much quicker in tensorflow so I think it's probably...

@wesselb This gets you pretty close! I get these losses for 5 epochs (the first data point is testing the untrained model). ``` Torch losses: [5.579, 5.314, 4.832, 4.471, 4.173,...

@wesselb [Here's](https://gist.github.com/DrJonnyT/c946044591fb4ce922b0f5e7fd0f047a) a slightly updated version that works for gnp and agnp. I tried convgnp but the part where it checks if the weights are the same fails, but the...

@wesselb Cool, I've made a minor tweak to that gist and now it works for convgnp as well 👍

@wesselb Sounds good!