diff options
Diffstat (limited to 'results/plain_cnn_p1_selector.json')
| -rw-r--r-- | results/plain_cnn_p1_selector.json | 366 |
1 files changed, 366 insertions, 0 deletions
diff --git a/results/plain_cnn_p1_selector.json b/results/plain_cnn_p1_selector.json new file mode 100644 index 0000000..8400f44 --- /dev/null +++ b/results/plain_cnn_p1_selector.json @@ -0,0 +1,366 @@ +{ + "expected_source": { + "author_crossover_commit": "90697e8a17ab5e0757a62df40da521766763abe1", + "main_commit": "dab65f657ed22ce90fe2b38a360ca62cc8b00e65", + "protocol_sha256": "345e0c1245f558a68ea2c212dfa427bdaa30eafd2c8d16c6763691ba354dae61" + }, + "gate": "pass", + "num_audited_records": 19, + "num_expected_records": 19, + "records": [ + { + "best_epoch": 10, + "best_validation_accuracy": 69.0199966430664, + "driver_wall_seconds": null, + "epochs_completed": 10, + "experiment_name": "plain-p1-bp-minicnn-lr0p025", + "final_validation_accuracy": 69.0199966430664, + "final_validation_loss": 0.9017932415008545, + "histogram_sha256": "534ebffb23256a16746d940493e8c7ea0d8c2685c97b364c11a855e653a6fce6", + "manifest": "runs/plain-p1-bp-minicnn-lr0p025/2026_07_27_13_34_41/crossover_manifest.json", + "method": "bp", + "rate": 0.025, + "test_metrics_all_nan": true, + "train_seconds": 24.938091278076172, + "validation_seconds": 2.485126495361328 + }, + { + "best_epoch": 3, + "best_validation_accuracy": 47.68000030517578, + "driver_wall_seconds": 53.98298001289368, + "epochs_completed": 10, + "experiment_name": "plain-p1-fa-minicnn-lr0p003", + "final_validation_accuracy": 42.47999954223633, + "final_validation_loss": 4.360357761383057, + "histogram_sha256": "0d2fff946e60672c21fe3789935e6d677198784b7405f30f8d0579e8e5a33e25", + "manifest": "runs/plain-p1-fa-minicnn-lr0p003/2026_07_27_13_35_54/crossover_manifest.json", + "method": "fa", + "rate": 0.003, + "test_metrics_all_nan": true, + "train_seconds": 37.1375572681427, + "validation_seconds": 2.595064878463745 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 44.02000045776367, + "driver_wall_seconds": 54.10479736328125, + "epochs_completed": 10, + "experiment_name": "plain-p1-fa-minicnn-lr0p01", + "final_validation_accuracy": 44.02000045776367, + "final_validation_loss": 243.5307159423828, + "histogram_sha256": "8117092fb45870c267bc1a36e37f98e2c4d05856a1f8c6a3de901014bcb50ae6", + "manifest": "runs/plain-p1-fa-minicnn-lr0p01/2026_07_27_13_36_48/crossover_manifest.json", + "method": "fa", + "rate": 0.01, + "test_metrics_all_nan": true, + "train_seconds": 37.05406403541565, + "validation_seconds": 2.487299680709839 + }, + { + "best_epoch": 8, + "best_validation_accuracy": 44.13999938964844, + "driver_wall_seconds": 53.24331092834473, + "epochs_completed": 10, + "experiment_name": "plain-p1-fa-minicnn-lr0p03", + "final_validation_accuracy": 42.08000183105469, + "final_validation_loss": 2647.7197265625, + "histogram_sha256": "85d5e24bc57523b004f45e4ed8180815701dbd8150708ec853bc4e6ab05fcc8d", + "manifest": "runs/plain-p1-fa-minicnn-lr0p03/2026_07_27_13_37_42/crossover_manifest.json", + "method": "fa", + "rate": 0.03, + "test_metrics_all_nan": true, + "train_seconds": 36.596099615097046, + "validation_seconds": 2.505472183227539 + }, + { + "best_epoch": 7, + "best_validation_accuracy": 50.099998474121094, + "driver_wall_seconds": 53.15471529960632, + "epochs_completed": 10, + "experiment_name": "plain-p1-dfa-minicnn-lr0p003", + "final_validation_accuracy": 44.52000045776367, + "final_validation_loss": 6.695240020751953, + "histogram_sha256": "a8a75822d50b6ff0a761bd80c6efb57a9c67744b81beea2162500579f8d33d7c", + "manifest": "runs/plain-p1-dfa-minicnn-lr0p003/2026_07_27_13_35_44/crossover_manifest.json", + "method": "dfa", + "rate": 0.003, + "test_metrics_all_nan": true, + "train_seconds": 36.10304522514343, + "validation_seconds": 2.478029489517212 + }, + { + "best_epoch": 7, + "best_validation_accuracy": 44.459999084472656, + "driver_wall_seconds": 53.555464029312134, + "epochs_completed": 10, + "experiment_name": "plain-p1-dfa-minicnn-lr0p01", + "final_validation_accuracy": 40.18000030517578, + "final_validation_loss": 207.70547485351562, + "histogram_sha256": "68c63c914a743afc85ca91f0a73ebf6ace618887117abcb6179dfac754ab8583", + "manifest": "runs/plain-p1-dfa-minicnn-lr0p01/2026_07_27_13_36_37/crossover_manifest.json", + "method": "dfa", + "rate": 0.01, + "test_metrics_all_nan": true, + "train_seconds": 36.20350170135498, + "validation_seconds": 2.5470991134643555 + }, + { + "best_epoch": 5, + "best_validation_accuracy": 37.68000030517578, + "driver_wall_seconds": 52.8665452003479, + "epochs_completed": 10, + "experiment_name": "plain-p1-dfa-minicnn-lr0p03", + "final_validation_accuracy": 28.860000610351562, + "final_validation_loss": 381436.09375, + "histogram_sha256": "cebe6d9d32ae33527173f83cd4d7c7478c8a4a53f11ed373f888a168c707af7b", + "manifest": "runs/plain-p1-dfa-minicnn-lr0p03/2026_07_27_13_37_31/crossover_manifest.json", + "method": "dfa", + "rate": 0.03, + "test_metrics_all_nan": true, + "train_seconds": 35.7019248008728, + "validation_seconds": 2.40809965133667 + }, + { + "best_epoch": 1, + "best_validation_accuracy": 27.479999542236328, + "driver_wall_seconds": 52.816155672073364, + "epochs_completed": 10, + "experiment_name": "plain-p1-pepita-minicnn-lr0p01", + "final_validation_accuracy": 10.880000114440918, + "final_validation_loss": 190006.296875, + "histogram_sha256": "e8e1c1ba5bab0ce023bfc7b5770b31bac3192003538b884d54eaeed2f7df6c84", + "manifest": "runs/plain-p1-pepita-minicnn-lr0p01/2026_07_27_13_31_55/crossover_manifest.json", + "method": "pepita", + "rate": 0.01, + "test_metrics_all_nan": true, + "train_seconds": 35.831242084503174, + "validation_seconds": 2.4864041805267334 + }, + { + "best_epoch": null, + "best_validation_accuracy": 4.86, + "driver_wall_seconds": 81.33231949806213, + "epochs_completed": 30, + "experiment_name": "plain-p1-ff-minicnn-lr0p03", + "final_validation_accuracy": 4.86, + "final_validation_loss": 2.253856658935547, + "histogram_sha256": "b4525d1e2cac272fb34dbc9e885d08c494b9b22a76e73abf726d365ac5d65f52", + "manifest": "runs/plain-p1-ff-minicnn-lr0p03/2026_07_27_13_30_34/crossover_manifest.json", + "method": "ff", + "rate": 0.03, + "test_metrics_all_nan": true, + "train_seconds": 64.50965976715088, + "validation_seconds": 3.194784641265869 + }, + { + "best_epoch": 1, + "best_validation_accuracy": 10.880000114440918, + "driver_wall_seconds": 144.76186728477478, + "epochs_completed": 10, + "experiment_name": "plain-p1-ep-minicnn-lr0p003", + "final_validation_accuracy": 10.880000114440918, + "final_validation_loss": 0.17824001610279083, + "histogram_sha256": "d9534ce6340c403bc3b32946d3327cdbba89e28a1d9edef708d6c6ea5b742fe5", + "manifest": "runs/plain-p1-ep-minicnn-lr0p003/2026_07_27_13_28_36/crossover_manifest.json", + "method": "ep", + "rate": 0.003, + "test_metrics_all_nan": true, + "train_seconds": 120.60259532928467, + "validation_seconds": 10.41963005065918 + }, + { + "best_epoch": 1, + "best_validation_accuracy": 10.880000114440918, + "driver_wall_seconds": 146.56798315048218, + "epochs_completed": 10, + "experiment_name": "plain-p1-ep-minicnn-lr0p01", + "final_validation_accuracy": 10.880000114440918, + "final_validation_loss": 0.17824001610279083, + "histogram_sha256": "fa782c1f8c6afbd6ee3f91d53d9fa799daf0c2b70356da12d7ae78440874903d", + "manifest": "runs/plain-p1-ep-minicnn-lr0p01/2026_07_27_13_31_01/crossover_manifest.json", + "method": "ep", + "rate": 0.01, + "test_metrics_all_nan": true, + "train_seconds": 121.54937195777893, + "validation_seconds": 10.896919965744019 + }, + { + "best_epoch": 1, + "best_validation_accuracy": 9.399999618530273, + "driver_wall_seconds": 146.23966145515442, + "epochs_completed": 10, + "experiment_name": "plain-p1-ep-minicnn-lr0p03", + "final_validation_accuracy": 9.399999618530273, + "final_validation_loss": 0.3396400213241577, + "histogram_sha256": "96edbf5cc1da7e44d57a8b130692aee42e20eb7c0964d01958b3ebdf1bd02b91", + "manifest": "runs/plain-p1-ep-minicnn-lr0p03/2026_07_27_13_33_28/crossover_manifest.json", + "method": "ep", + "rate": 0.03, + "test_metrics_all_nan": true, + "train_seconds": 121.5877935886383, + "validation_seconds": 10.519296884536743 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 69.18000030517578, + "driver_wall_seconds": 117.89603066444397, + "epochs_completed": 10, + "experiment_name": "plain-p1-dualprop-minicnn-lr0p025", + "final_validation_accuracy": 69.18000030517578, + "final_validation_loss": 0.9032290577888489, + "histogram_sha256": "0aa63a064b4e8db582b406edbbac232d927c84a8b60aba5d554457cf1c0eda37", + "manifest": "runs/plain-p1-dualprop-minicnn-lr0p025/2026_07_27_13_28_36/crossover_manifest.json", + "method": "dualprop", + "rate": 0.025, + "test_metrics_all_nan": true, + "train_seconds": 101.47104024887085, + "validation_seconds": 2.532742738723755 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 71.31999969482422, + "driver_wall_seconds": 63.29983401298523, + "epochs_completed": 10, + "experiment_name": "plain-p1-clean_kp-minicnn-lr0p003", + "final_validation_accuracy": 71.31999969482422, + "final_validation_loss": 0.8342198729515076, + "histogram_sha256": "2e56f287295009cc55f7efc35bd111aa80f846f85b6bf8bf8caa53aed984ed3c", + "manifest": "runs/plain-p1-clean_kp-minicnn-lr0p003/2026_07_27_13_38_35/crossover_manifest.json", + "method": "clean_kp", + "rate": 0.003, + "test_metrics_all_nan": true, + "train_seconds": 46.737804889678955, + "validation_seconds": 2.5253994464874268 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 72.62000274658203, + "driver_wall_seconds": 63.30412268638611, + "epochs_completed": 10, + "experiment_name": "plain-p1-clean_kp-minicnn-lr0p01", + "final_validation_accuracy": 72.62000274658203, + "final_validation_loss": 0.8119363188743591, + "histogram_sha256": "8c78f13ab1e650aa05e0387f47b213c2d765462d59ce00806136fcecc100409c", + "manifest": "runs/plain-p1-clean_kp-minicnn-lr0p01/2026_07_27_13_39_39/crossover_manifest.json", + "method": "clean_kp", + "rate": 0.01, + "test_metrics_all_nan": true, + "train_seconds": 46.380820989608765, + "validation_seconds": 2.5903303623199463 + }, + { + "best_epoch": 9, + "best_validation_accuracy": 66.77999877929688, + "driver_wall_seconds": 62.69386029243469, + "epochs_completed": 10, + "experiment_name": "plain-p1-clean_kp-minicnn-lr0p03", + "final_validation_accuracy": 66.73999786376953, + "final_validation_loss": 0.9755602478981018, + "histogram_sha256": "fda1bde32375de1afd20390609112587af959d6e046b8ecfeba31539a3ce42e0", + "manifest": "runs/plain-p1-clean_kp-minicnn-lr0p03/2026_07_27_13_40_42/crossover_manifest.json", + "method": "clean_kp", + "rate": 0.03, + "test_metrics_all_nan": true, + "train_seconds": 45.855462312698364, + "validation_seconds": 2.5258145332336426 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 71.18000030517578, + "driver_wall_seconds": 75.57752275466919, + "epochs_completed": 10, + "experiment_name": "plain-p1-sdil-minicnn-lr0p003", + "final_validation_accuracy": 71.18000030517578, + "final_validation_loss": 0.8337415456771851, + "histogram_sha256": "1ab9ecfef74354866e3071bd4911bf2e776ff6a4493b14a76364c02ad090af19", + "manifest": "runs/plain-p1-sdil-minicnn-lr0p003/2026_07_27_13_38_23/crossover_manifest.json", + "method": "sdil", + "rate": 0.003, + "test_metrics_all_nan": true, + "train_seconds": 51.720041275024414, + "validation_seconds": 2.5472254753112793 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 72.5999984741211, + "driver_wall_seconds": 75.49908399581909, + "epochs_completed": 10, + "experiment_name": "plain-p1-sdil-minicnn-lr0p01", + "final_validation_accuracy": 72.5999984741211, + "final_validation_loss": 0.8222209215164185, + "histogram_sha256": "b1d425eb285b9ab367f851df8a3c103a85ef81bf44a6ba12d47edba955866215", + "manifest": "runs/plain-p1-sdil-minicnn-lr0p01/2026_07_27_13_39_39/crossover_manifest.json", + "method": "sdil", + "rate": 0.01, + "test_metrics_all_nan": true, + "train_seconds": 52.18812847137451, + "validation_seconds": 2.575517416000366 + }, + { + "best_epoch": 10, + "best_validation_accuracy": 66.5999984741211, + "driver_wall_seconds": 74.3778088092804, + "epochs_completed": 10, + "experiment_name": "plain-p1-sdil-minicnn-lr0p03", + "final_validation_accuracy": 66.5999984741211, + "final_validation_loss": 0.9670084118843079, + "histogram_sha256": "b420ad1f13c38e4eb469cf228aa65692f21ae97e3d0e3c6b9119c5f3b36ef492", + "manifest": "runs/plain-p1-sdil-minicnn-lr0p03/2026_07_27_13_41_44/crossover_manifest.json", + "method": "sdil", + "rate": 0.03, + "test_metrics_all_nan": true, + "train_seconds": 51.435088872909546, + "validation_seconds": 2.528696298599243 + } + ], + "selected": { + "bp": { + "best_validation_accuracy": 69.0199966430664, + "experiment_name": "plain-p1-bp-minicnn-lr0p025", + "rate": 0.025 + }, + "clean_kp": { + "best_validation_accuracy": 72.62000274658203, + "experiment_name": "plain-p1-clean_kp-minicnn-lr0p01", + "rate": 0.01 + }, + "dfa": { + "best_validation_accuracy": 50.099998474121094, + "experiment_name": "plain-p1-dfa-minicnn-lr0p003", + "rate": 0.003 + }, + "dualprop": { + "best_validation_accuracy": 69.18000030517578, + "experiment_name": "plain-p1-dualprop-minicnn-lr0p025", + "rate": 0.025 + }, + "ep": { + "best_validation_accuracy": 10.880000114440918, + "experiment_name": "plain-p1-ep-minicnn-lr0p003", + "rate": 0.003 + }, + "fa": { + "best_validation_accuracy": 47.68000030517578, + "experiment_name": "plain-p1-fa-minicnn-lr0p003", + "rate": 0.003 + }, + "ff": { + "best_validation_accuracy": 4.86, + "experiment_name": "plain-p1-ff-minicnn-lr0p03", + "rate": 0.03 + }, + "pepita": { + "best_validation_accuracy": 27.479999542236328, + "experiment_name": "plain-p1-pepita-minicnn-lr0p01", + "rate": 0.01 + }, + "sdil": { + "best_validation_accuracy": 72.5999984741211, + "experiment_name": "plain-p1-sdil-minicnn-lr0p01", + "rate": 0.01 + } + }, + "selection_rule": "maximum best validation accuracy, then finite final validation loss, then lower learning rate", + "stage": "plain_cnn_p1", + "test_policy": "none" +} |
