stage2/adaptformer/selection.json

{
  "schema_version": 1,
  "study": "lr_capacity_search",
  "method": "adaptformer",
  "protocol": "frame_block_cv_search",
  "model_seed": 42,
  "split_seed": 42,
  "init_from_sha256": "3094f103c17bd13558f960c22b91ed3316679cadeab88ba269d862019c4dd58a",
  "fold_subset": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9
  ],
  "smoke": false,
  "search_space": {
    "lr_candidates": [
      0.0001,
      0.0003,
      0.001
    ],
    "capacity_axis": "bottleneck",
    "reference_capacity": {
      "bottleneck": 64
    },
    "capacity_candidates": [
      1,
      16,
      32,
      64
    ]
  },
  "lr_selection": {
    "axis": "lr",
    "candidates": [
      0.0001,
      0.0003,
      0.001
    ],
    "held_at": {
      "bottleneck": 64,
      "reduction_factor": 12
    },
    "fold_subset": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7,
      8,
      9
    ],
    "mean_accuracy_by_candidate": [
      {
        "lr": 0.0001,
        "config_id": "adaptformer__lr1.0e-04__bottleneck0064",
        "mean_accuracy": 0.958181818181818
      },
      {
        "lr": 0.0003,
        "config_id": "adaptformer__lr3.0e-04__bottleneck0064",
        "mean_accuracy": 0.9727272727272727
      },
      {
        "lr": 0.001,
        "config_id": "adaptformer__lr1.0e-03__bottleneck0064",
        "mean_accuracy": 0.9845454545454546
      }
    ],
    "highest_mean_accuracy": 0.9845454545454546,
    "tied_highest": [
      0.001
    ],
    "tie_rule": "highest ten-fold mean held-out accuracy, then the numerically lowest learning rate on an exact tie. Accuracy alone: no UAR, no weighted F1, no loss and no tolerance band enters it",
    "selected": 0.001
  },
  "capacity_selection": {
    "axis": "bottleneck",
    "candidates": [
      1,
      16,
      32,
      64
    ],
    "at_lr": 0.001,
    "fold_subset": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7,
      8,
      9
    ],
    "mean_accuracy_by_candidate": [
      {
        "bottleneck": 1,
        "config_id": "adaptformer__lr1.0e-03__bottleneck0001",
        "mean_accuracy": 0.9654545454545455,
        "trainable_params": 32274,
        "reduction_factor": 768
      },
      {
        "bottleneck": 16,
        "config_id": "adaptformer__lr1.0e-03__bottleneck0016",
        "mean_accuracy": 0.9818181818181818,
        "trainable_params": 308934,
        "reduction_factor": 48
      },
      {
        "bottleneck": 32,
        "config_id": "adaptformer__lr1.0e-03__bottleneck0032",
        "mean_accuracy": 0.9736363636363636,
        "trainable_params": 604038,
        "reduction_factor": 24
      },
      {
        "bottleneck": 64,
        "config_id": "adaptformer__lr1.0e-03__bottleneck0064",
        "mean_accuracy": 0.9845454545454546,
        "trainable_params": 1194246,
        "reduction_factor": 12
      }
    ],
    "highest_mean_accuracy": 0.9845454545454546,
    "tied_highest": [
      64
    ],
    "tie_rule": "highest ten-fold mean held-out accuracy, then fewer trainable parameters, then the lexicographically smallest configuration id. Accuracy alone decides first: no UAR, no weighted F1, no loss and no tolerance band enters it",
    "selected": 64
  },
  "selected": {
    "config_id": "adaptformer__lr1.0e-03__bottleneck0064",
    "values": {
      "bottleneck": 64,
      "lr": 0.001,
      "reduction_factor": 12
    },
    "trainable_params": 1194246,
    "mean_accuracy": 0.9845454545454546,
    "cell_dir": "outputs/search__adaptformer__src-ferplus__seed42/cells/adaptformer__lr1.0e-03__bottleneck0064"
  },
  "optimism": "the configuration is selected on the same ten held-out folds whose mean accuracy is then reported, so the comparison is optimistically biased by hyperparameter selection on top of the best-epoch-on-the-held-out-block optimism every cell already carries. It is not nested cross-validation, not independent validation, not an unbiased estimate and not a like-for-like comparison with the published figures",
  "design": "a sequential hyperparameter study: learning-rate selection, then capacity selection at the selected rate. The best-observed configuration of each strategy is the configuration that enters the five-method comparison",
  "comparison_eligible": true,
  "comparison_eligibility": "selected over all ten folds"
}