SDMLab

Srinivas Deekonda Machine Learning Lab — a from-scratch machine learning / neural-network library, built incrementally as a long-term learning and engineering project.

What this is

SDMLab implements the core building blocks of ML/deep-learning frameworks — tensors, autograd, layers, optimizers, and eventually classic ML algorithms — from first principles in Python, instead of wrapping PyTorch, TensorFlow, or JAX. The goal is to understand how these frameworks actually work internally by building them, one concept at a time, rather than to compete with any of them.

Priority order: understanding and correctness first, API convenience second.

Philosophy

Every component follows the same loop before it's considered done:

learn → implement → test → document → release

Development happens in small, regular iterations. Each concept — a layer, an activation, an optimizer — gets implemented, tested, documented, and shipped as its own version bump rather than bundled into one large release. See Versioning below for how that's tracked.

Current scope: neural networks, from scratch

• Tensors with automatic differentiation — computational graphs, backpropagation
• Dense/Linear layers, activation functions, loss functions
• Optimizers: SGD, Momentum, AdaGrad, RMSProp, Adam, learning-rate scheduling, weight decay
• Batch / mini-batch training, a Sequential model abstraction, training loops, metrics
• Dataset utilities
• A backend abstraction over NumPy (CPU) today, CuPy (GPU) later, switched with sdmlab.set_backend("numpy" | "cupy")

Classic ML (linear/logistic regression, SVMs, decision trees, clustering, ...) and other paradigms are explicitly planned for later — the architecture is designed so they slot in as new modules without requiring a rewrite of anything above. See docs/ARCHITECTURE.md for the full folder structure and the reasoning behind it, and docs/nn/CODING_GUIDE.md for the coding conventions used while building the neural-network module.

Example (target API)

from sdmlab.layers import Dense
from sdmlab.activations import ReLU
from sdmlab.losses import CrossEntropy
from sdmlab.optimizers import Adam

model = sdmlab.Sequential([
    Dense(784, 128),
    ReLU(),
    Dense(128, 10),
])
optimizer = Adam(learning_rate=0.001)

This is the API being built toward — check the version history below for what's actually implemented at any given point.

Installation

pip install sdmlab

The package name is registered on PyPI; the library itself is early-stage and under active development, so check the installed version against Versioning before relying on any particular API surface.

Versioning

Two independent things are versioned here:

• Releases follow semantic versioning (MAJOR.MINOR.PATCH), tracked in CHANGELOG.md and sdmlab.__version__. A MINOR bump generally corresponds to one new concept landing (a layer, an optimizer, ...). To depend on a specific release, pin it the normal pip way: pip install sdmlab==0.1.0.
• The public API is versioned separately as a namespace: sdmlab.v1, and later sdmlab.v2 if a MAJOR bump ever requires one. Plain import sdmlab always follows whichever namespace is currently the default — that's a deliberate choice made in this repo, not something you opt into — while code that wants to freeze against one API shape forever can import sdmlab.v1 directly. Full mechanism: docs/VERSIONING.md.

Project layout

src/sdmlab/       the package
tests/            mirrors src/sdmlab/ 1:1
examples/         one runnable script per milestone
docs/             architecture notes and per-module coding guides

Full details: docs/ARCHITECTURE.md.

License

GNU General Public License v3.0

Author

Srinivas Deekonda