GSoC 2025

 Finished up Test I (mostly). Now onto Specific Test II - Lens Finding. Downloaded the data... wow, way more non-lenses than lenses. This class imbalance is serious! Accuracy won't mean much here. Need to focus on ROC AUC and maybe precision/recall for the lens class. Sticking with ConvNeXt V2 but need to add class weighting to the loss function. This feels more like a real-world astro problem.

 After some experimentation, fine-tuning ConvNeXt V2 Tiny seems to be the way to go for Test I. Setting up the custom dataset loader, handling the 1-channel to 3-channel conversion, applying ImageNet normalization... It's training! Watching the validation AUC climb is oddly satisfying. Need to calculate the ROC/AUC properly for the report.

 Okay, time to tackle the GSoC tests. Downloaded the "Common Test I" dataset. Three classes: no substructure, sphere, vortex. Looks like a classic image classification problem. The Deep Learning courses are paying off! Decided to go with PyTorch and try fine-tuning a pre-trained model. Maybe a ResNet? Or something newer... ConvNeXt V2 looks promising. Let the coding commence!

 Finally making some progress interacting with a mock LSST repository using the Butler. butler.get('calexp', ...) It works! Retrieving calibrated exposures, checking metadata... It's complex, abstracting away the file system, but I can see how powerful it is for managing petabytes of data. Still feels a bit like black magic though.

Reached the Convolutional Neural Networks (CNNs) course in the Deep Learning Specialization. YES! This feels directly relevant. Processing images... detecting features... This is exactly what lens finding and classification models do. Suddenly the DeepLense project tasks make a lot more sense. Getting excited about actually applying this. 

While wrestling with the LSST setup, also started Andrew Ng's Deep Learning Specialization. Course 1: Neural Networks and Deep Learning. Feels like a good refresher and goes deeper than the ML one. Vectorization, activation functions... Good stuff. Need this foundation solid if I'm going to build a pipeline for deep learning models. 

 Alright, time to get the LSST stack running locally. Following the lsstinstall guide... dependencies... conda conflicts... Docker maybe? Spent a good chunk of the day just getting the basic LSST commands to run without throwing errors. Small victory, but a victory nonetheless! One step at a time.

 Trying to wrap my head around this project. LSST Science Pipelines, the Butler API... it's a whole ecosystem. And DeepLense – classifying dark matter substructure? Super-resolution? Feeling a mix of "this is awesome" and "where do I even start?". Time to hit the docs. Hard.

Okay, GSoC orgs and projects announced! Scrolled through... ML4Sci... DeepLense... "Data Processing Pipeline for the LSST". Huh. Rubin Observatory, massive data, deep learning for lens finding... This sounds intense. And exactly the kind of astro+coding challenge I was looking for. Reading the description..