Structured, benchmarked Python solutions to Deep-ML practice problems — every problem explained from first principles and implemented across multiple frameworks, from raw Python up to hand-written CUDA kernels.
Deep-ML is a LeetCode-style platform for machine learning and deep learning fundamentals — linear algebra, probability, neural network internals, and classical ML algorithms implemented from scratch.
This repository is my solutions log, built to a consistent standard: every problem gets a plain-language explanation of the underlying math, a working implementation, and a written walkthrough of why the solution works — not just code that passes the test case. Where a problem is a good excuse to go deeper (e.g. a core linear algebra op), solutions are also implemented across the frameworks that actually run production ML systems: NumPy, PyTorch, raw CUDA, and Tinygrad.
Each problem lives in its own folder, named after the problem, containing exactly three files:
Deep-ML_Problems/
├── <problem>/
│ ├── README.md
│ ├── solution.py
│ └── explanation.md
│
└── README.md # you are here
| File | Purpose |
|---|---|
README.md |
The problem itself — restated clearly, with the exact function signature, a worked example, and constraints, plus a link back to the original Deep-ML problem page. |
solution.py |
A correct, self-tested Python implementation. Simpler problems use one clear approach; foundational ones (e.g. core tensor/linear-algebra ops) include multiple implementations for comparison. |
explanation.md |
The reasoning: the underlying math, a step-by-step trace on the example input, complexity analysis, and — where relevant — a comparison of approaches and when to reach for each one. |
| # | Problem | Category | Difficulty | Frameworks | Solution |
|---|---|---|---|---|---|
| 1 | Matrix-Vector Dot Product | Linear Algebra | Easy | Python · NumPy · PyTorch · CUDA · Tinygrad | Problem 01-Matrix_Vector Dot Product |
| 2 | Transpose of a Matrix | Linear Algebra | Easy | Python · NumPy · PyTorch · CUDA · Tinygrad | Problem 02-Transpose of a Matrix |
New problems are added as they're solved — this table is the single source of truth for progress. See Roadmap for what's next.
| Layer | Tools |
|---|---|
| Language | Python 3.10+ |
| Numerical computing | NumPy |
| Deep learning frameworks | PyTorch, Tinygrad |
| GPU programming | CUDA (via PyCUDA), for problems where a hand-written kernel is instructive |
| Testing | Lightweight self-test blocks per solution (if __name__ == "__main__":), asserting output against the problem's stated example(s) and edge cases |
git clone https://github.com/Ayush-2703/Deep-ML_Problems.git
cd Deep-ML_Problems/<problem-folder>
pip install -r requirements.txt # numpy at minimum; torch/tinygrad for framework variants
python solution.py # runs the self-test blockEach solution.py is runnable standalone and prints pass/fail for every test case it checks itself against — no separate test runner needed.
Most "LeetCode solutions" repos are a wall of code with no context — useful for nobody, including future-me. The three-file structure here is deliberate:
README.mdmakes each folder self-contained — you don't need the Deep-ML site open to understand the problem.explanation.mdis the part most solution repos skip, and the part that actually matters: why the approach works, not just that it does.- Multi-framework solutions, where included, exist because the same operation (e.g. a matrix-vector product) looks completely different depending on whether you're calling BLAS, letting an autograd engine trace it, writing the GPU kernel yourself, or watching a minimal framework compile one for you — and seeing all four side by side is the fastest way to actually understand what's happening under
nn.Linear().
- Work through Deep-ML's Linear Algebra track
- Work through the Deep Learning / Neural Networks track
- Work through the Probability & Statistics track
- Add a progress badge / completion percentage to this README
Distributed under the MIT License. See LICENSE for details.
You're free to use, fork, and build on this for personal and commercial projects.
If this repository helped you, please consider giving it a ⭐
It takes 2 seconds and helps others discover it.