AmeerSayyad/smai-a4-numpy-mlp

SMAI Assignment 4: selected MLP (Q1.2.5 and Q1.5)

3072-64-10 ReLU MLP (197,322 parameters), CIFAR-10, seed 1040.

• numpy_mlp_selected.npz / .json: the NumPy MLP selected in 1.2.5 (stratified 3600/1200/1200 split, lr 0.01, batch 64, early stopping): validation accuracy 0.4881, test accuracy 0.4918.
• torch_mlp_p0.npz / torch_models_info.json: the PyTorch MLP compared with the IOC-NN in 1.5 (50,000-image train part, true labels, 100 epochs): test accuracy 0.4529. The IOC-NN is in AmeerSayyad/smai-a4-ioc-nn.
• Weights: W{i} shape (n_in, n_out), b{i} shape (n_out,). Preprocessing: flatten to 3072, divide by 255, standardise each column with the mean/std of the respective training part.