Pranav-Sriv/smai-assignment-4-ioc

SMAI Assignment 4: IOC-NN, clean labels

Final epoch-100 model compared in Section 1.5 of the assignment. Trained from scratch on CIFAR-10's original 50,000 training images, with the original 10,000 test images used for reporting. Architecture was selected in Section 1.2.5 using validation accuracy; no test-based architecture selection was performed for this experiment.

Architecture: 3072 -> 64 -> 10, ELU hidden activation, Identity output logits, 197,322 parameters. Training used mean cross-entropy, SGD, learning rate 0.05, batch size 64, one batch per update, 100 epochs, no early stopping, seed 11023, and no label noise.

Final train accuracy: 63.630%; test accuracy: 40.750%; mean test F1: 0.406282. These are single-seed assignment results and are not a production benchmark.

Files

• model.pt: PyTorch state dictionary, without optimizer or credentials.
• weights.npz: equivalent weights and biases, keys W0, b0, W1, b1; each W has shape (input_width, output_width).
• preprocessing.npz: per-column mean and standard deviation fitted on the 50,000 training images after division by 255; zero standard deviations are replaced by 1.
• config.json: architecture, training settings, class order, preprocessing information, and final metrics.

Load locally

After downloading this folder, run:

import numpy as np
import torch
from torch import nn

model = nn.Sequential(nn.Linear(3072, 64), nn.ELU(),
                      nn.Linear(64, 10), nn.Identity())
model.load_state_dict(torch.load("model.pt", map_location="cpu", weights_only=True))
model.eval()
statistics = np.load("preprocessing.npz", allow_pickle=False)
mean = torch.tensor(statistics["mean"], dtype=torch.float32)
std = torch.tensor(statistics["std"], dtype=torch.float32)

# images: uint8 NumPy array with shape (N, 32, 32, 3), RGB.
def predict(images):
    x = torch.from_numpy(images.reshape(-1, 3072)).float()
    x = (x / 255.0 - mean) / std
    with torch.no_grad():
        return model(x).argmax(dim=1)

Flatten in HWC row-major order; do not transpose to CHW. Outputs are logits, and the predicted class is their argmax. Class names are recorded in config.json.

The IOC-NN uses ELU and constrains every linear weight after the first layer, including the output layer, to be non-negative. After each SGD step, negative constrained weights were replaced by exp(w - 5). Its initial first-layer parameters matched the paired MLP.