saimuralipr/smai-a4-mlp

SMAI A4: Multi-Layer Perceptron (ReLU 128-64)

This model repository contains the serialized parameters (.npz format) and architecture specifications for SMAI Assignment 4 (Monsoon 2026) at IIIT Hyderabad.

Model Architecture

• Input Dimension: $3072$ ($32 \times 32 \times 3$ flattened CIFAR-10 RGB pixels)
• Hidden Layer 1: $3072 \to 128$ (Linear + ReLU)
• Hidden Layer 2: $128 \to 64$ (Linear + ReLU)
• Output Layer: $64 \to 10$ (Linear + Softmax / Cross-Entropy Loss)
• Total Parameters: $402,250$

Preprocessing Information

• Standardization: Channel-wise mean and standard deviation computed strictly on the 36,000 training images and applied identically to test data.
  • Channel Means: [0.4914, 0.4822, 0.4465] (normalized to $[0, 1]$ before scaling)
  • Channel Stds: [0.2470, 0.2435, 0.2616]
• Dataset: CIFAR-10 (36,000 train / 12,000 val / 12,000 test stratified split)

Performance

• Validation Accuracy: $48.35%$
• Test Accuracy: $47.73%$ (NumPy) / $47.17%$ (PyTorch)
• Macro F1-Score: $0.4749$

Loading Instructions

The weights can be loaded using NumPy or the course repository's Model.load_from(path) API:

import numpy as np

data = np.load("MLP_ReLU_128_64.npz")
for k in data.files:
    print(k, data[k].shape)