Invictus22/smai-a4-cifar10-mlp

SMAI A4 — CIFAR-10 MLP

• Architecture: 3072-128-10 (hidden activation: relu, output: 10 logits, no activation)
• Trained on the CIFAR-10 train part (50,000 images, true labels, p = 0) for 100 epochs with plain SGD: lr = 0.01, batch size = 64, grad_accumulation_steps = 1, cross-entropy loss, seed = 14015.
• Ordinary MLP, weights of any sign.
• Preprocessing: flatten the 32x32x3 uint8 image to 3072 values (HWC order), divide by 255, then subtract preprocess_mean and divide by preprocess_std (both stored in model.npz, computed on the train part).
• model.npz: W0, b0, W1, b1, ... with W of shape (in, out), so z = h @ W + b; activations lists the activation of every layer.
• Test accuracy (10,000 test images): 0.4776