Ayyuce Demirbas

Saved my Linux kernel configuration for future use via
“cp .config kernel-6.6-$(date +%Y%m%d).config”

github.com/ayyucedemirb...

Compiling the Linux kernel, pretraining my image feature extractor using contrastive learning (on a remote GPU via AnyDesk), and working on my paper.

Happy Star Trek day! Live long and prosper. 🖖🏻 #StarTrek #StarTrekDay

Training a multimodal model for 200 epochs. This should be done in about a week. Wish me luck. 🤞🏻

Instead of treating molecular features as nodes, I made patients the nodes. Now similar patients share information through graph connections, learning from each other's outcomes. Check out the implementation: github.com/ayyucedemirb...

I’ve implemented suffix arrays and suffix trees from scratch to find repeats, perform fast pattern matching, and generate the Burrows-Wheeler Transform. It also creates visualizations of the data structures. Check out the code on GitHub: github.com/ayyucedemirb...

Local alignment with Smith-Waterman algorithm

github.com/ayyucedemirb...

Global alignment with Needleman-Wunsch algorithm

github.com/ayyucedemirb...

This is how you can create a longest common subsequence graph using dynamic programming.

github.com/ayyucedemirb...

A GNN project to classify BRCA patient survival using multi-omics data (RNA-seq, CNV, mutations, proteins).

github.com/ayyucedemirb...

Now you can link your ORCID account to your GitHub account.

Predicting cancer subtypes using multi-omics data with Graph Convolutional Networks. I know the results are not good enough but I’m working on it. 🧬
GitHub repo: github.com/ayyucedemirb...

The Ultimate Guide to Fine-Tuning LLMs from Basics to Breakthroughs

arxiv.org/pdf/2408.13296

You can use the method below to compute BLEU‑1 through BLEU‑4, METEOR, ROUGE‑1/2/L, and BERTScore metrics on your validation set at the end of each epoch.

If you are getting a warning that says, "A decoder-only architecture is being used, but right-padding was detected! For correct generation results, please set padding_side='left' when initializing the tokenizer." you should add this to your code: processor.tokenizer.padding_side = 'left'.

A long time ago in a galaxy far, far away…
May the 4th be with you. ✨⚔️

Wow, that’s some amazing progress. An epoch takes an hour and the validation loss reduces by only 0.00022.

There has been a recurring issue with Kaggle notebooks for the past few days, the session freezes while training models. You don't even know which epoch this will happen in, and sometimes hours of training are lost because of this freeze.

This weekend, 📝I worked on my IEEE conference paper, 📚 wrote a blog post on Cache-Augmented Generation: ayyucedemirbas.github.io/posts/cag.ht..., and 👩🏻‍💻solved 9 LeetCode problems: leetcode.com/u/ayyuce.
🎶 Listened to Puccini's Tosca while working: open.spotify.com/album/1muIgA...
Productive weekend.😇

GitHub - ayyucedemirbas/RAG_techniques Contribute to ayyucedemirbas/RAG_techniques development by creating an account on GitHub.

Made a multimodal retrieval-augmented generation example handling images and text.
👩🏻‍💻 Code: github.com/ayyucedemirb...
🤗 Hugging Face Demo: huggingface.co/spaces/ayyuc...

I’m trying to fine-tune the blip-vqa-base model that consists of 385M parameters on my visual instruction tuning dataset named vqa-rad-instructions. Since the dataset contains only 1793 samples for training, I’m using k-fold to prevent overfitting.

Mathematical optimization: finding minima of functions

scipy-lectures.org/advanced/mat...

You can use these commands to list the sizes of all your conda environments:

for env in $(conda env list | awk '{print $1}' | grep -v "#"); do echo $env; du -sh $(conda env list | grep "$env" | awk '{print $NF}'); done

conda env list | grep -o '/.*' | while read env; do du -sh "$env"; done

Now you can visualize folder structures with:
- JSON/TXT exports (--output) for automation
- Custom truncation (--truncate N) to control file listings
Try it:
pip install --upgrade PeekDir
PeekDir your_folder --output structure.json --format json --truncate 8

To print your computer’s memory size you can use these:

On a Linux distro: grep MemTotal /proc/meminfo | awk '{print $2 / 1024 / 1024 " GB"}'

On macOS: sysctl hw.memsize | awk '{printf "%.2f GB\n", $2 / (1024*1024*1024)}'

That’s a great idea, thank you so much. I’ll work on that asap.

Just released PeekDir, a Python tool to visualize directory structures. It's useful for understanding the folder structures of complex datasets.

Installation: pip install PeekDir
Usage: PeekDir [directory]

Example usage and output below! 👇

GitHub - ayyucedemirbas/SSD_multimodal_sensor_fusion Contribute to ayyucedemirbas/SSD_multimodal_sensor_fusion development by creating an account on GitHub.

SSD sensor fusion. Used Bird’s-Eye View Projection (LiDAR point clouds can be projected onto a top-down view, generating a pseudo-image that encodes spatial density, height, and intensity) and converted radar data into 2D maps that capture range and velocity information.:

github.com/ayyucedemirb...

NeoProtein-GPT is a specialized text-generation model for designing novel protein sequences.

GitHub: github.com/ayyucedemirb...
Model: huggingface.co/ayyuce/NeoPr...
Hugging Face Space: huggingface.co/spaces/ayyuc...

GitHub - ayyucedemirbas/GPU_programming Contribute to ayyucedemirbas/GPU_programming development by creating an account on GitHub.

I’m learning CUDA GPU programming basics. You can find the examples on this repo. Since my computer has an Apple silicon (M1) I’m using Colab. I share both the scripts (.cu and .py files) and Jupyter notebook files of these examples. I use both C++ and Python (pyCUDA). github.com/ayyucedemirb...