poojith06/smai-a4-mlp

SMAI A4 - MLP (3072-256-128-10)

SMAI Assignment 4 (Monsoon 2026), Q1. Final-epoch model (not a checkpoint) trained with true labels (p = 0) for 100 epochs.

• Architecture: 3072 - 256 - 128 - 10, hidden activation ReLU, output = 10 logits, 820,874 parameters
• Training: plain SGD, learning rate 0.05, batch size 64, CCE loss, seed 22010, no early stopping
• Preprocessing: divide pixel values by 255 (float64), then (x - mean) / std for each of the 3072 columns. The mean and std are in preprocessing_train50k.npz (keys mean, std).
• Weights: .npz with W0, b0, W1, b1, ..., z = x @ W + b, W of shape (n_in, n_out).
• Results (final epoch): train accuracy 0.9076, test accuracy 0.4877

The file numpy_m3_weights.npz is the NumPy MLP selected in Task 1.2.5 (test accuracy 0.4957), saved with Model.save_to; its preprocessing is in preprocessing_split36k.npz (mean and std of the 36,000-image training split).

See config.json for the full configuration.