dharmask/smai-a4-mlp

MLP CIFAR-10 final clean-label model

Architecture: 3072-64-32-10; float64; ten logits. Final epoch: 100.

Files: state_dict.pt (PyTorch), weights.npz (NumPy row-vector weights), config.json (architecture, labels, seeds and training metadata), preprocessing.npz (official-training-only mean/std/scale), inference.py.

Minimal inference, from this folder (NumPy required):

from inference import predict
classes, logits = predict(raw_uint8_rgb_minibatch, '.')

PyTorch reload in the assignment repository, only for Q1.4:

import json, torch
from src.q1.pytorch_section14.models import fresh_model
cfg = json.load(open('config.json', encoding='utf-8'))
model = fresh_model(cfg['kind'], cfg['seed'], cfg['widths'])
model.load_state_dict(torch.load('state_dict.pt', map_location='cpu', weights_only=True))
model.eval()

Apply the saved preprocessing before a direct PyTorch forward pass. Softmax is optional after logits; the IOC convexity guarantee is for logits.