smai-a4-iocnn

Input-Output Convex Neural Network (IOC-NN) trained on CIFAR-10 for SMAI Assignment 4 (Q1, Section 1.4/1.5).

• Architecture: 3072 -> 256 (ELU) -> 10 (logits)
• Initialisation: IOC-NN: first layer = MLP init, layers >= 2 uniform(0, 2/n_in); after every step negative weights w of layers >= 2 -> exp(w - 5)
• Training: SGD, lr 0.02, batch size 64, 100 epochs, true labels, seed 1146
• Preprocessing: flatten the 32x32x3 image (HWC order) to 3072 values, divide by 255, standardise with mean/std from preprocessing.npz (computed on the 50,000 training images)
• Files: model.npz (W0, b0, W1, b1; W has shape n_in x n_out), preprocessing.npz, config.json

Load the arrays with np.load("model.npz"); the forward pass is logits = elu(x @ W0 + b0) @ W1 + b1.