Skip to content

mlscratch v0.2.0

Latest

Choose a tag to compare

@Mattral Mattral released this 19 Jun 09:47
· 19 commits to main since this release

mlscratch v0.2.0 — Major Expansion

This release delivers the first full expansion of the library, adding neural networks, supervised learning algorithms, metrics, preprocessing utilities, and runnable examples.

Added

  • Neural module (mlscratch.neural):
    Perceptrons, MLPs, Autoencoders, RNN/LSTM, CNN building blocks, Transformer stack, GAN, Hopfield, RBM, RBF, Complex‑valued NN — all with tests.
  • Supervised module (mlscratch.supervised):
    Linear models and KNN implementations, with test coverage.
  • Bayesian module finalized:
    Naive Bayes, Gaussian Process, HMM, Kalman Filter, Bayesian NN/Network — with tests.
  • Reinforcement module finalized:
    QLearning, DQN, DDPG, TD3, PPO, SAC — plus utilities and tests.
  • Unsupervised module expanded:
    DBSCAN, PCA, GMM, ICA, KMedoids, Apriori, t‑SNE, AgglomerativeClustering — with tests.
  • Metrics: accuracy, precision, recall, F1, confusion matrix, MSE, R² — with tests.
  • Preprocessing: StandardScaler, MinMaxScaler, train_test_split — with tests.
  • Examples: runnable scripts per module demonstrating usage.

Infrastructure

  • Version bump to 0.2.0 in pyproject.toml and package __init__.py.
  • Updated CHANGELOG.md with neural and supervised entries.
  • CI workflow extended to cover new modules.
  • Documentation updated (docs/index.md, mkdocs.yml, roadmap).

Tests

  • 700+ tests total across all modules.
  • Three‑tier structure per algorithm: API · Correctness · Edge Cases.
  • Analytically verifiable assertions (Bellman update, d‑separation, Kalman smoother monotonicity, GPR interpolation, etc.).

This release establishes mlscratch as a comprehensive from‑scratch ML library, spanning unsupervised, supervised, Bayesian, reinforcement, and neural methods, with metrics, preprocessing, examples, and full test coverage.