Shardul007/cifar10-mlp-p0

MLP (ReLU)

CIFAR-10 classifier from SMAI Assignment 4 (Q1), trained on the clean labels (noise p = 0) for 100 epochs.

• Architecture: 3072 -> 64 -> 32 -> 10, ReLU hidden activations, identity output (logits), 199,082 parameters
• All weights unconstrained; He-normal initialisation, biases 0
• Loss: categorical cross-entropy; plain SGD, lr 0.05, batch size 64, no gradient accumulation, no early stopping
• Seed: 1077 (torch.manual_seed(1078), DataLoader generator seed 1079)
• Test accuracy (CIFAR-10 test part, 10,000 images): 0.4478

Preprocessing

Flatten the 32x32x3 image to 3072 values, convert to float64, divide by 255, then standardise each column: x = (x / 255 - mean) / std with mean and std taken from preprocessing.npz (computed on the 50,000 CIFAR-10 train images; columns with std 0 use divisor 1).

Files

• mlp_weights.pt - PyTorch state_dict of the nn.Sequential model (Linear/activation pairs, last layer Linear only)
• mlp.npz - same weights for the NumPy library of the assignment (W0..W2 of shape (n_in, n_out), b0..b2; load with Model.load_from)
• preprocessing.npz - mean, std (3072 values each)