https://github.com/ercodex/predicting-new-medicines-with-sgd-classifier

"Shittiest Model of the Competition" award winner!
https://github.com/ercodex/predicting-new-medicines-with-sgd-classifier

classifier-model dask kaggle-competition machine-learning numpy python scikit-learn

Last synced: 5 months ago
JSON representation

"Shittiest Model of the Competition" award winner!

• Host: GitHub
• URL: https://github.com/ercodex/predicting-new-medicines-with-sgd-classifier
• Owner: ercodex
• License: mit
• Created: 2025-01-26T13:29:20.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2025-01-26T14:31:37.000Z (over 1 year ago)
• Last Synced: 2025-02-03T02:06:00.858Z (over 1 year ago)
• Topics: classifier-model, dask, kaggle-competition, machine-learning, numpy, python, scikit-learn
• Language: Python
• Homepage:
• Size: 31.3 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# *If you’re looking for a story of triumph, you’re in the wrong place. If you want to see how grit and curiosity can turn failures into lessons, welcome aboard!*

# "Shittiest Model of the Competition" Award Winner!

- This model was my submission for the Kaggle competition **"NeurIPS 2024 - Predict New Medicines with BELKA"**.

## Story

- Fresh off completing a machine learning course on Udemy, I thought, "Why not test my skills?" So I jumped into this competition... solo. Spoiler alert: the DL veterans crushed me.  

- Here's what I had: 8 GB of DDR4 RAM, a trusty GTX 1650 GPU, minimal ML experience, zero knowledge of neural networks, and, of course, my unmatched stubbornness.

## What I Did

1. **Understanding the Data**  

   - The dataset was biological and massive: 3 protein targets, millions of molecule sequences generated by ~300 molecules. The task? Predict which molecules are likely to bind to these targets. Cool concept, overwhelming data.  

2. **The 50 GB Challenge**  

   - Yep, 50 GB of data. My RAM said, "Nope." I tried compressing, transforming, mapping... eventually shrunk it to 35 GB. Still not enough.  

   - A college instructor gave me a golden tip: "If they can zip the download file to 5 GB for competitors, you can figure something out." Inspired, I spent **2 months** creating mapping tables, using the less-space-consuming data types , and trimming down data to a manageable 15 GB. Turns out, others in the discussion forum were struggling too.

3. **Building the Model**  

   - With the clock ticking, I went for the two most practical option for me: **Random Forest Classifier**, or **Stochastic Gradient Descent Classifier**. No fancy hyperparameter tuning, there was barely any time.

   - Training posed another hustle. The dataset was too large for my setup, and NumPy just didn’t cut it. Enter **Dask library**, the lifesaver! It handled chunked training like a charm, and I managed to train, predict, and submit just before the deadline.

   - Each training session nearly took 8 hours. My laptop was working like a full-time employee.

4. **The Results**  

   - Final ranking: **1925/1952**. Sure, I wasn’t at the top, but the journey taught me invaluable lessons in data handling, persistence, and problem-solving. Plus, it was hilarious!

5. **Special Thanks**

   - A shoutout to my instructor who guide the way to the model. **Not turning away an eager student can change many things.**  

   - A shoutout to **ChatGPT**, which sped up my problem-solving process. While I didn’t use it to code the model, it served as a trusted guide whenever I got stuck.

## What I Learned

- Learned techniques to handle **large datasets**.  

- Discovered the **Dask** library.

- Learned about **SGD Classifier**

- Gained hands-on experience with .csv and .parquet formats.

- SMILES format for representing biological data.

- Gained practical experience in **data visualization** and **feature engineering**.  

- Realized I need to start training myself in **Deep Learning**.  

- Built connections with instructors by showcasing my efforts.  

- Most importantly, I had fun and grew as a Computer Engineering student.

**There is a picture about what the name of training data looked like at the end. It's pretty funny! Check it out.**