Inquiry on GSP ID (1,) embedding

[kwon-encored]

@Sukh-P

Hello! I hope this message finds you well!

After successfully trying out your architecture on my side-project data, I was able to obtain favorable results. Thank you to your team!

During my attempt, I did not incorporate your GSP embedding as described:

Before being fed into the 3D convolutional layers, the satellite and two NWP inputs were appended with an additional channel to hold a learned embedding for the GSP ID.

Questions

1. Code Location for GSP Embedding:
   Unlike the most updated version of PVNet, I was imitating your journal paper version (back in 2024). Within your current repository, where can I find the related code for embedding the GSP to each weather data?

2. Dimension Clarification:
   In your paper, Figure 4 shows "GSP" embed(16,). Where does that 16 come from?

3. Handling Solar Coordinates:
   From your paper version, does the (16,2) solar coordinate go through a fully connected layer, or does the raw data directly go into cat_mod with the other 1D encoded datasets after flattening?

   Once the NWP and satellite data have been encoded into 1D vectors, they are concatenated along with the calculated solar coordinates, and another embedding of the GSP ID.

[AUdaltsova]

Hi @kwon-encored! Amazing to hear you've got some results, please do let us know if/when you put your work somewhere we can access! Also massive well done, reusing our code is not easy currently, it's great to hear that it worked out for you in the end. (Also if you want do let me know what ended up solving your issues, I'm really curious - but no pressure!)

1. I think the relevant bits would be in this file, here in the init and here in the forward run. I'm pretty sure this part was not affected by our recent migration from ocf_datapipes to ocf-data-sampler, but if in doubt look at the version of PVNet before 4.0 (this is when it starts being ocf_datapipes incompatible).
2. This is the dimension of the embedding, it can be set in the configs with embedding_dim: 16. As far as I know, the number was chosen empirically.
3. So here in the forward pass you can see that sun features go through a fc with out_features = 16, which is how this dimensionality stays constant - sun position itself will have the same length as the time window (see here), which is not always 16!

Hope this helps, let me know if you have further questions!