yusiwen/dl-from-scratch

feat(ml): add GBDT and random forest implementations

feat: add hyperparameter grid search wrapper

feat: add CNN vs ViT benchmark script (untrained)

feat: add CV inference demos (vit/demo.py, unet/demo.py, yolo/demo.py)

refactor: restructure into domain-grounded directories (ml/ cv/ gen/ graph/ rl/ nlp/)

move: lora/ → nlp/lora/

feat: add LoRA (parameter-efficient fine-tuning for GPT)

feat: add t-SNE (nonlinear dimensionality reduction with PCA comparison)

feat: add MobileNet (depthwise separable convolutions)

feat: add GCN, DQN, SimCLR, YOLO (4 new models + notebooks)

feat: add DDPM (Denoising Diffusion Probabilistic Model)

feat: add Seq2Seq Transformer (encoder-decoder with cross-attention)

feat: add VAE (Variational Autoencoder)

feat: add ResNet50 (Bottleneck block)

docs: update ROADMAP with new candidate models and training estimates

fix: regenerate notebooks with get_device() from utils.device

refactor: extract device detection to utils/device.py (CUDA→MPS→CPU)

feat: add P2+P3 notebooks (basics 9, Word2Vec, LSTM)

feat: add P1 notebooks for CNN, BERT, ResNet18, ResNet34

chore: ignore .ipynb_checkpoints/

chore: add jupyterlab + ipykernel, register kernel

feat: add Jupyter notebooks for GPT, ViT, DCGAN, UNet

docs: add Notebooks section to ROADMAP

docs(roadmap): restructure completed items and update future directions

fix: remove duplicate cnn/ entry in README project tree

feat: add DCGAN, ViT, and UNet modules

feat(resnet): add resnet34 and rename resnet to resnet18

infra: add config system, TensorBoard logging, and reproducibility seed

refactor: GPT word-level tokenizer, full text8, bad token blocking

feat: add GPT (Decoder-only Transformer + Causal Attention + KV Cache)

docs: update BERT accuracy to realistic 50% (character-level limitation)

feat: add LSTM from scratch (hand-written gates) for IMDB sentiment

refactor: BERT uses HuggingFace datasets instead of embedded texts

docs: add Core Concepts section mapping models to key ML/DL concepts

feat: add Word2Vec (CBOW + Skip-gram + Negative Sampling)

feat: add BERT (Transformer Encoder + MLM) for NLP

docs: add PCA, k-NN, Perceptron to Models table, fix resnet model note

feat: add PCA, k-NN, and Perceptron

feat: add Decision Tree (ID3, Iris) and Naive Bayes (Gaussian, MNIST)

docs: fix typo Plato -> Platt SMO

feat: add SVM (GD primal + SMO dual) with Linear and RBF kernels

chore: stop tracking model weights (.pt .npz) in git

feat: add Linear Regression (Normal Equation + GD) on California Housing

feat: add basics/logistic_regression and basics/k_means

feat: add pure NumPy MLP for MNIST (97.9% test acc)

docs: add comprehensive English comments to resnet/ scripts

feat: complete CNN training (82.4% test acc on CIFAR-10)

docs: update README with CNN info, HF datasets description, remove old data/ dir from tree

refactor: migrate to HuggingFace datasets library

feat: add SimpleCNN for CIFAR-10 classification