Closed yellowcap closed 5 months ago
I suggest we make embeddings of a subset of the training data location, and compare. Probably both the usual average embeddings (at the patch level) and the smaller ones at the self-attention patch level.
In the latter case, we would only have those corresponding to the groups we used as inputs, right?
So we compare:
Input data | Embedding | Case |
---|---|---|
All 13 bands | Normal ones (average at patch level 5kmx5km ) |
The normal ones |
All 13 bands | self-attention patch groups 300mx300m |
Fine-level embeddings |
Only RGB | Normal ones (average at patch level 5kmx5km ) |
Restricted input, same output |
Only RGB | self-attention patch groups | Fine embedding of band group 300mx300m |
After some discussions, we realized that there is a technical possiblity to run the Clay v0 model with partial inputs. We need to