shubhamLearnsMachine

A learning memoir of my AI ML journey — small scripts and projects as I work through concepts, not a polished library. Each subfolder is a self-contained topic.

Contents

Folder	What's in it
pytorch-basics/	Small scripts covering PyTorch tensor fundamentals — creation, numpy interop, device placement, indexing/broadcasting, and autograd.
mnist_pytorch_mlp/	A custom MLP (nn.Module) trained on MNIST, built step-by-step: manual training loop → reusable training function → comparing Full-batch vs Mini-batch vs Stochastic Gradient Descent on convergence speed and overfitting risk.
mnist-numpy-ann/	A neural network built from scratch in raw NumPy (forward pass, backprop, gradient descent — no autograd frameworks) trained on MNIST digit classification. Includes the debugging process and training results.
cnn/my_conv2d/	A from-scratch Conv2D layer (MyConv2d, im2col + matmul) verified for correctness/gradients against nn.Conv2d and benchmarked CPU vs GPU across problem sizes. See cnn/my_conv2d/README.md for the full write-up and findings.
cnn/cifar10/	A CNN trained on CIFAR-10, improved step-by-step from 73.86% → 82.43% test accuracy purely through training methodology — data augmentation, batch norm, best-checkpoint restoration, patience-based early stopping, and a random hyperparameter search — with confusion-matrix/misclassified-image analysis at the end. See cnn/cifar10/README.md for the full timeline and findings.

Setup

Environment used: a conda env (mlwork) with numpy, matplotlib, torch, and keras (Keras is only used for MNIST data loading / one-hot encoding, not for building or training any model here).