DivyG007/smai-a4

SMAI Assignment 4 — CIFAR-10 MLP and IOC-NN (Q1.5)

The two models compared in Section 1.5, both trained on the CIFAR-10 HF train part (50,000 images, true labels) for 100 epochs with plain SGD (lr 0.01, batch size 64), seed 14009.

Folder	Model	Layers	Hidden activation	Parameters	Test accuracy
cifar10-mlp/	MLP	3072-128-10	ReLU	394,634	0.4819
cifar10-ioc-nn/	IOC-NN (weights of layers ≥ 2 non-negative)	3072-128-10	ELU	394,634	0.4847

Each folder has model.npz (arrays W0, b0, W1, b1; W{i} has shape (n_in, n_out), output = h @ W + b; load with Model.load_from from mlp.ipynb), preprocessing.npz (mean, std of the 3072 input columns), config.json (architecture, training and preprocessing) and a short README.md.

Preprocessing: flatten the 32×32×3 image (HWC order) to 3072 values, divide by 255, subtract mean and divide by std.