Запуск через скрипты¶
Один эксперимент:
from cnn_neat import ExperimentConfig, run_single_experiment
cfg = ExperimentConfig(positive_class=3, num_generations=50, population_size=100)
run_single_experiment(cfg)
Или JSON (пример со всеми дефолтами — configs/experiment.defaults.json):
Исторический протокол 05¶
pop=1000, A→B→C→D, без конуса размножения и кэша хэшей.
Рабочий протокол «один против остальных»¶
python scripts/ablation/06_ova_class_study.py # 10 классов, C/D + 5-fold CV
python scripts/ablation/09_ova_structural_hash_study.py
python scripts/ablation/10_ova_hash_host_spill_study.py # planned cache; full_gpu | cpu_mut_fwd_gpu_gd
python scripts/ablation/11_ova_gpu_only_study.py # CUDA + cone (основной)
python scripts/ablation/11_ova_gpu_only_study.py --start-phase c
python scripts/ablation/12_ova_cpu_ram_ab_study.py # Phase A без кэша
Отложенная проверка (holdout) чемпионов: скрипты 11b / 11c в scripts/ablation/.
MNIST, CIFAR-10 и многоклассовый режим (скрипты 13–20)¶
python scripts/ablation/15_ova_mnist_gpu_only_study.py # MNIST OVA A-D (script 11 twin)
python scripts/ablation/13_merge_ova_multiclass_seed.py # merge 10 OVA champions -> seed
python scripts/ablation/14_multiclass_evolve.py --seed-dir runs/ablation/ova_multiclass_seed # evolve, no GD
python scripts/ablation/14b_multiclass_evolve_gd.py --seed-dir runs/ablation/ova_multiclass_seed # evolve + periodic GD
python scripts/ablation/16_mnist_multiclass_from_ova.py --generations 100 --resume # MNIST multiclass queue
python scripts/ablation/17_mnist_per_node_act_study.py # MNIST per-node σ OVA A/B + merge queue
python scripts/ablation/18_cifar_per_node_act_study.py # CIFAR analog of 17; val=20k, pop=25
python scripts/ablation/19_mnist_val5k_light_gd_study.py # like 17, val=5k, light GD, apply cap
python scripts/ablation/20_dynamic_pop_cache_study.py # occupancy pop + genome stagn + dual-track GD; drop cache if occupancy <10 → grow ×1.2 from current (ceiling 2048)
Протоколы: Multiclass из OVA, MNIST OVA, per-node activations (17), CIFAR per-node activations (18), val=5k light GD (19).
Инфографики:
python scripts/analysis/export_script11_ab_final_infographic.py
python scripts/analysis/export_script11_c_final_infographic.py
python scripts/analysis/export_cache_capacity_infographic.py
python docs/build_site_figures.py
Прочее: scripts/ablation/02_arch_ablation.py, scripts/run_cifar10_class_suite.py, scripts/bench/benchmark_forward_epoch.py.