rohday/smai-a4-mlp

SMAI A4 Q1: MLP on CIFAR-10

{
  "model": "MLP",
  "architecture": [
    3072,
    128,
    10
  ],
  "hidden_activation": "ReLU",
  "output": "10 logits (identity), softmax cross-entropy loss",
  "weights_format": "weights.npz: W{i} of shape (n_in, n_out) and b{i}, layer i computes h @ W{i} + b{i}",
  "preprocessing": "flatten to 3072 values (32 x 32 x 3, HWC order), divide by 255, subtract mean and divide by std from preprocessing.npz (computed on the 50000 cifar-10 training images)",
  "training": {
    "labels": "true labels (p = 0)",
    "epochs": 100,
    "lr": 0.02,
    "batch_size": 64,
    "grad_accumulation_steps": 1,
    "optimiser": "plain SGD",
    "constraint": null
  },
  "test_accuracy": 0.4725
}