Results after training the model with the masking strategy described in the previous post. In particular, I mask out 17x17 rectangles. This is squared EMD as before.
Things are getting much more interesting:
| Strategy | Samples |
|---|---|
| Greedy |
|
| Antigreedy |
|
| Orderless |
|
| Gibbs |
|
It’s definitely looking like I need to tweak the Gibbs parameters. I’ll also do another hyperparameter search to make sure I’ve got the right architecture.