deep-learning-from-scratch

A complete neural network built entirely in x86 assembly language that learns to recognize handwritten digits from the MNIST dataset. No frameworks, no high-level languages - just pure assembly - ~5.3× faster than NumPy

🔬 The most atomic GPT-2 implementation in 265 lines of pure Python & CUDA. A bilingual "Rosetta Stone" for understanding LLM internals from scratch. No dependencies, just math and kernels.

Deep Learning Framework only using numpy: Linear, Convolution, Flatten, Max and Mean Pooling layers, activation functions, Softmax, MSE and Cross Entropy. Useful functions: train, save an load a model to deploy it, get nb of parameters, draw learning curves

Single-layer Perceptron (Logistic Regression) binary classifier in Python, powered by a high-performance C++20 matrix engine (matrixlib) via Pybind11.

📖 ⟪밑바닥부터 시작하는 딥러닝2⟫ 스터디 기록 저장소

Vision Transformer (ViT) image classifier written from scratch in PyTorch: train on CPU/GPU/multi-GPU, predict with attention heatmaps, Streamlit demo, Kaggle/Colab notebook, 86 tests. 97% on synthetic shapes, 86% on PlantVillage plant diseases.

Deep learning from scratch in pure NumPy — tensors, layers, losses, and optimizers with hand-written backward passes and no autograd. A learning project with math notes for every algorithm.

A zero-dependency Dense MLP neural network built completely from scratch in NumPy, featuring custom activation overflow clipping and modular data pipelines.

A pure-NumPy deep learning framework built from scratch. Features a custom reverse-mode autograd engine, VJPs, causal multi-head attention, and AdamW.

A lightweight, deep learning library written in pure Python

A coding exercises with numpy and pytest for Deep Learning From Scratch

Deep Learning from scratch: mathematical derivations, manual forward/backward implementations of MLPs, CNNs, RNNs, LSTMs and GRUs, followed by practical PyTorch implementations of modern neural network architectures. Written in Python & PyTorch.

"A from-scratch implementation of a modern Large Language Model (LLaMA architecture). Features full backpropagation calculus in pure NumPy, followed by a GPU-accelerated PyTorch training loop."

A Python project implementing a neural network framework from scratch using NumPy. Includes fully connected (Dense) layers, ReLU and Sigmoid activations, a simple SGD optimizer, and a minimal training loop. Designed for hands-on learning of neural network fundamentals without relying on any deep learning frameworks

A transformer up to Mixture-of-Experts, trained in NumPy with every gradient derived by hand and checked by finite differences. RoPE, GQA, SwiGLU, RMSNorm, KV-cache, int8, SFT. 196 tests, one script reproduces every result.

从零实现的自动微分引擎(200 行, 批量节点 + 广播 + 拓扑反向传播), 用来训练 MLP 并实时可视化决策边界与损失曲线。41 项无依赖测试含数值梯度校验。

Industrial corrosion rate predictor implemented from scratch in pure NumPy using electrochemical Nernst-Planck equations.

A from-scratch neural network built in NumPy to classify handwritten digits (MNIST), demonstrating forward propagation, backpropagation, and gradient descent without deep learning frameworks.

밑바닥부터 시작하는 딥러닝 스터디 요약 저장소 (HW)

Final Year Project (FYP) - An AI-powered examination attendance system with handwritten text recognition.