LearnCV.ai

Teaching AI visually and practically!

---

Project maintained by purus15987 Hosted on GitHub Pages — Theme by mattgraham

🧪 Practice Lab – 2: NumPy & ML Model Design

🎯 Objective:

Practice building and testing basic machine learning algorithms from scratch using only NumPy (no TensorFlow/PyTorch for model internals). Also practice clean coding with OOP (Object-Oriented Programming).

---

💻 Lab Problems (Choose one)

Easy Level

1. Linear Regression – do both:

   • Closed-form solution (normal equation)
   • Gradient Descent (add Ridge regularization)

2. Data Split & Evaluation – create functions to:

   • Split data into train/validation/test sets
   • Apply scaling
   • Calculate metrics (MSE, MAE, R²)

---

Medium Level

1. Logistic Regression (Binary):

   • L2 regularization
   • Stable math using log-sum-exp
   • SGD or mini-batch gradient descent

2. K-Nearest Neighbours (KNN):

   • Different distance metrics
   • Weighted voting

3. Small Feed-Forward Neural Network:

   • 1–3 hidden layers
   • Activation functions (ReLU, Sigmoid, Tanh)
   • Batch normalization option

---

Hard Level

1. CNN Basics:

   • Implement conv2d and maxpool functions
   • Train a small CNN on part of MNIST

2. Simple Autoencoder:

   • Reduce dimensions
   • Show reconstruction quality

3. Optimization Algorithms:

   • Implement and compare: SGD, Momentum, RMSProp, Adam
   • Show and analyze convergence plots

---

An Engineer’s Trek into Machine Learning

source

---

🛠️ Lab Instructions

1. Pick one problem from above.

2. Plan your code before starting — draw a small diagram of your components:

   • Example: DataLoader → Model → Optimizer → Trainer → Metrics
3. Write Python + NumPy code (vectorized, avoid loops over each sample).

4. You can use matplotlib for plots, and scikit-learn only for:

   • Loading datasets
   • Checking results (not for internal model work)

5. Create unit tests to check:

   • Correct shapes
   • Gradients (numeric check)
   • Loss decreases during training

6. Write a short report (2–4 pages, PDF):

   • Problem statement
   • Design choices
   • Algorithms used
   • Experiments + plots
   • Key results and conclusions
7. Submit your code + tests + report in a ZIP or GitHub repo.

---

📂 Suggested Project Structure

README.md          # Instructions to run
dataset/           # dataset files (.csv / .png / .xsl)
results/           # experimentation results 
src/               # Code files
  |_model.py
  |_train.py
  |_utils.py
  |_dataloader.py
  |_main.py
  |_eval.py
  |_test.py
tests/             # Unit tests
report.pdf         # Your short write-up

---

📊 Example: Linear Regression

Components:

• DataLoader → loads dataset
• LinearModel → predict, loss, gradient
• Optimizer → SGD with learning rate
• Trainer → training loop, early stopping

Tests:

• Shapes match
• Loss decreases each epoch
• Gradient error < 1e-5

---

🏆 Marks Distribution (Total 35)

Criteria	Marks
Understanding problem	5
Correct, vectorized code	10
Modular design, OOP	7
Experiments & analysis	6
Tests & reproducibility	4
Report clarity	3
Total	35

---

🙏 Credits / Acknowledgements

We acknowledge the use of open-source educational content and inspiration from the following sources:

• NumPy Documentation – https://numpy.org/doc/
• Matplotlib Documentation – https://matplotlib.org/stable/contents.html
• Python Data Science Handbook by Jake VanderPlas (O’Reilly, open-access version available online)
• scikit-learn Documentation – https://scikit-learn.org/
• An Engineer’s Trek into Machine Learning – www.ml4devs.com
• CS231n: Convolutional Neural Networks for Visual Recognition – http://cs231n.stanford.edu/
• Dive into Deep Learning (D2L.ai) – https://d2l.ai/
• Machine Learning Mastery – Code Algorithm from Scratch – https://machinelearningmastery.com/

---

This lab is intended solely for academic and practice purposes.

Full credit goes to the original authors and maintainers for their contribution to the open-source and data science education community.

---