Christina (Chrisy)

Day 51: Looking at my neural network layers with deep dream! #100DaysOfCode

thank youuu!! 🚀

Day 50: Dataloader for Kandinsky patterns is done! Now I need to find a loss term that makes my network not only sparse but also modular! So in love with my PhD team, they help me verbalise my problems 🤓 #100DaysOfCode

Day 49: Meeting with my Doctoral supervisor and advisors was fun as always! Great new ideas for how to proceed - mainly dataset and loss term related. #100DaysOfCode

Day 48: Optimisation of graph visualisation in 3D!! (currently not working) #100DaysOfCode

Day 47: Doing some optimisation. It's working! at least in a small scale! #100DaysOfCode

Day 46: Goal to get all of the umap plots look similar: failed. Tried with transform once, fit multiple times. Maybe the amount of datapoints is the problem. Keep going, keep growing! #100DaysOfCode

Day 45: Yesterday, I tried to further understand my data. The labels don't make too much sense to me, hence messaged a clinician! #100DaysOfCode

Day 44: Plotting data for genes related to AMD. Super excited to feed a neutral network with it combined with images 😍 #100DaysOfCode

Day 43: Music to my ears (or to my laptops ears) - new cuda version after 4 years! #100DaysOfCode

Day 42: Suddenly everything stopped working ... #100DaysOfCode

Day 42: Yesterday I was trying to implement a dataloader and for whatever magic reason it's not working. #100DaysOfCode

Some cool prof in eye research just asked me what I published. Well, there is this project on an idea for a mission to Neptune in a journal. I have a feeling he wanted another answer 👀 but well, I do what I can 😂

Python code

Day 41: Another day of #100DaysOfCode! I fixed a problem with gradients. Really just switched to manual optimisation in PyTorch lightning, which resulted in the need of changing the callback for saving checkpoints. Seems to work now. Next thing - troubleshooting the visualisation (I don't trust).

Day 40: More coding for modular training!! #100DaysOfCode

Model training

Day 39: All bugs fixed! (this will probably jinx it, but oh well). Model is training! #100DaysOfCode

I know, ended up just needing another .unsqueeze luckily! Right now I am using random public datasets, but yes, eventually satellite images with bee colony loss labels.

TabPFN Code

Secondly, I have been dissecting some scikit-model code.

Python Error Message

Day 38: Two things today!! Number one - problems with dimensions for the metrics. Have to translate them back from one-hot I guess. #100DaysOfCode

Day 37: Baseline testing! #100DaysOfCode

Modal split in Vienna keeps trending in the right direction. In 2024 the utilization for day-to-day distances:

34% public transport
30% walking
25% car
11% bike

Cars have never been this low, bikes never this high.

Saw my first ever and second rocket launches in the past 24h!!! Super amazing experience, and the sunrise was the cherry on top 😭😍

and the actual update - trying to translate pytorch into torch lightning!

Day 36: My torch lightning attempts to gracefully shut down! Maybe I should do the same with my coding career. [joke] #100DaysOfCode

Day 35: Lots of modularity and honey bee colony loss write-up! #100DaysOfCode

Day 34: Made progress on my code today thanks to @mateosuffern.bsky.social! Turning some pytorch code into pytorch lightning. My favourite part are the sanity checks! #100DaysOfCode

Cartoon noting the irony that car infrastructure is commonly considered a public investment, and public transit infrastructure is often attacked as a wasteful subsidy, when the opposite is actually true.

"Public investment" vs "Wasteful Subsidy." The only problem with this clever Singer cartoon is that some people might actually not get that it’s sarcastically illustrating the incorrect perception problem, NOT telling the truth.

Just in case it actually needs to be said, the truth is the opposite.

Python/Matplotlib, showing (bad) prediction results.

Day 33: Clearly not working. Still, I made progress with understanding LSTM. #100DaysOfCode

Error messages and way too high loss values.

Day 32: I promise, I am an expert in deep learning. #100DaysOfCode

Day 31: Searching for some benchmarking datasets and metrics! #100DaysOfCode