train model + evaluate model

[peterdudfield]

clear script to train model + evaluate model. We might have this already, but we should have clear docs for this.

Train the ML model. We want to follow on from #6

[siddharth7113]

Hi @peterdudfield

This issue looks like a great place for me to jump in! I have some experience with ML/DL projects from college, but I’m excited to learn more about how these tools are applied in a real-world sceanrios. I’ve already read through the getting_started.md guide. Could you point me toward any particular issue for getting started? I’m looking forward to contributing and learning!

[peterdudfield]

Thats great @siddharth7113 , thanks for this. I think the first thing to do is #8 and #1 . This will make samples ready for PVnet to be trained

[Sharkyii]

Hey @peterdudfield ,

I’m Sneh, a new contributor, and I’d be excited to take on this issue. With #1 and #8 now completed, this feels like the natural next step in the pipeline.

Proposed plan:

• Review the existing training and evaluation scripts to understand the current flow and identify missing pieces.
• Build a clean, modular training pipeline that aligns with the outputs produced so far.
• Define a clear evaluation process using standard metrics to support downstream benchmarking.
• Add concise documentation so contributors can easily run and reproduce results.

If there are any preferred patterns, tools, or design choices you’d like this to follow, I’d be happy to align with them. Looking forward to contributing further! 🚀
Could you please assign this issue to me?

[peterdudfield]

I think that would be great @Shark-Bot-X, @siddharth7113 can you share you GSOC report, just to help with these things.

[siddharth7113]

Sure @peterdudfield , Here is my gsoc-report link.

@Shark-Bot-X A good place to get upto speed is to look at the getting started docs here

Our current plan is to make a similar model configuration with US data and porting to new version of ocf-data-sampler, I will try to open issues related to it in few days time, meanwhile you can look into the repository here

[Sharkyii]

Hey @siddharth7113 ,
Thank you for sharing the report!

I had a few doubts. I’ve gone through the ocf-data-sampler, pvnet, and open-data-pvnet repositories, and I’m trying to understand where I should begin contributing.

Would you recommend that I start by adapting the pipeline for the US datasets (I don’t have much prior knowledge in this area and plan to contribute to it later), or should I first focus on building the training and evaluation scripts for the UK datasets, since those are already merged and more stable?

[Sharkyii]

Hi @peterdudfield @siddharth7113 ,
I’m having some difficulty obtaining the GSP data in Zarr format. I’ve searched through the repository but couldn’t find clear instructions on how to generate or download it. Every attempt so far results in an access-related error, indicating that I may not have the required permissions.

It would be really helpful if you could point me to the exact process or location for downloading or generating the GSP data in Zarr format.

Quick update on my progress: I’ve successfully built the training pipeline and am currently testing it. I’m still encountering a few runtime errors, which I’m in the process of debugging.

Thanks in advance for your help.

[peterdudfield]

Hey

You should be able to get it from s3://ocf-open-data-pvnet/data/uk/pvlive/v2/combined_2023_gsp.zarr/, and if this is not documentated, we should definately add that.

@siddharth7113 might have a more update to date version though

[Sharkyii]

@peterdudfield Thanks for helping, it’s working now. The issue turned out not to be with the GSP data itself, but with the configuration parameters on my end, specifically include_sun and include_time. Once those were corrected, i will start working on the evaluation pipeline and then submit a PR.