arkagme/smai-a4-mlp

MLP (relu) on CIFAR-10 (SMAI Assignment 4, Q1.4/1.5, clean labels p = 0)

• Architecture: 3072 -> 128 -> 10; hidden activation relu, output: identity (10 logits), loss: softmax cross-entropy.
• Initialisation: N(0, 2/n_in) for ReLU (1/n_in otherwise), biases 0.
• Training: plain SGD, lr 0.01, batch size 64, grad accumulation 1, 100 epochs, float64, seed 1044, on the 50,000-image CIFAR-10 train split (true labels).
• Result: final train accuracy 0.8611, test accuracy 0.4624 (10,000-image test split).
• Preprocessing: image (32, 32, 3) uint8 -> reshape(3072) (H, W, C order) -> float64 / 255 -> (x - mean) / std with mean, std from preprocessing.npz (per-column statistics of the 50k training split, population std).
• Files: model.npz holds W0, b0, W1, b1, ... with W of shape (n_in, n_out) so that logits = h @ W + b, plus activations and loss.