File size: 2,584 Bytes
2f6b87a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 | {
"schema_version": 1,
"study": "lr_capacity_search",
"method": "linear_probe",
"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": [
0.0003,
0.001,
0.003
],
"capacity_axis": null,
"reference_capacity": {},
"capacity_candidates": null
},
"lr_selection": {
"axis": "lr",
"candidates": [
0.0003,
0.001,
0.003
],
"held_at": {},
"fold_subset": [
0,
1,
2,
3,
4,
5,
6,
7,
8,
9
],
"mean_accuracy_by_candidate": [
{
"lr": 0.0003,
"config_id": "linear_probe__lr3.0e-04",
"mean_accuracy": 0.889090909090909
},
{
"lr": 0.001,
"config_id": "linear_probe__lr1.0e-03",
"mean_accuracy": 0.9227272727272726
},
{
"lr": 0.003,
"config_id": "linear_probe__lr3.0e-03",
"mean_accuracy": 0.9327272727272726
}
],
"highest_mean_accuracy": 0.9327272727272726,
"tied_highest": [
0.003
],
"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": 0.003
},
"capacity_selection": null,
"selected": {
"config_id": "linear_probe__lr3.0e-03",
"values": {
"lr": 0.003
},
"trainable_params": 4614,
"mean_accuracy": 0.9327272727272726,
"cell_dir": "outputs/search__linear_probe__src-ferplus__seed42/cells/linear_probe__lr3.0e-03"
},
"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"
}
|