jung-geun/PSO
1
0
Fork 0
mirror of https://github.com/jung-geun/PSO.git synced 2026-09-20 14:11:48 +09:00
Code Issues Packages Projects Releases Wiki Activity
Files
master
PSO/benchmark_results/pso_v6_heavy_tasks.json
T
jung-geun 813433000a feat: modernize PSO and add convergence research 
Migrate the package and examples to the tensor-native PyTorch implementation, add benchmark evidence, and add the guarded post-training convergence protocol with TensorBoard progress monitoring and hash-verified recovery.

Constraint: Preserve one-shot official-test sealing and auditable research artifacts

Rejected: Commit local .omc runs and downloaded datasets | multi-gigabyte runtime state is machine-local

Confidence: high

Scope-risk: broad

Not-tested: Production CUDA run on pieroot-server
2026-09-07 22:03:25 +09:00

2080 lines
60 KiB
JSON
Raw Permalink Blame History

{
"protocol_version": "HEAVY-TASK-PSO-V6 1.0.0",
"timestamp": "2026-09-02T14:52:52Z",
"hardware": {
"platform": "macOS-26.6.2-arm64-arm-64bit",
"system": "Darwin",
"machine": "arm64",
"processor": "arm",
"python_version": "3.11.15",
"torch_version": "2.13.0",
"pso_version": "4.0.0",
"device_type": "mps",
"device_str": "mps",
"mps_available": true
},
"pso_version": "4.0.0",
"official_test_data_loaded": false,
"official_test_evaluations": 0,
"study_design": {
"screen": {
"methods": [
"G0",
"G5",
"G6",
"G8"
],
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"cells": 16
},
"confirmation": {
"methods_per_workload": {
"mnist_compact": [
"G8",
"G6"
],
"mnist_wide": [
"G8",
"G5"
],
"fashion_compact": [
"G8",
"G6"
],
"fashion_wide": [
"G8",
"G5"
]
},
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
]
},
"selection": {
"candidates": [
"G0",
"G5",
"G6"
],
"ranking": [
"validation_nll_ascending",
"validation_accuracy_descending"
],
"confirmation_reference": "G8"
},
"feasibility_contract": {
"execution": "all_confirmed_runs_complete_with_finite_metrics",
"minimum_validation_nll_reduction_fraction": 0.2,
"minimum_validation_accuracy_gain_percentage_points": 20.0,
"requires_both_optimization_thresholds": true
},
"validation": {
"source": "official_training_split_only",
"split": "50000_search_10000_validation_stratified",
"split_seed": 20260902,
"official_test_split_used": false
}
},
"method_configs": {
"G0": {
"config_id": "G0",
"scale_type": "per_tensor_sd",
"init_position_mode": "antithetic",
"position_radius": 0.5,
"initial_velocity_radius": 0.0,
"mutation_prob": 0.0,
"reset_velocity_radius": 0.02,
"reflective_bound": 3.0,
"projection_seed": null,
"latent_dim": "full",
"description": "exact V5 control"
},
"G5": {
"config_id": "G5",
"scale_type": "per_tensor_sd",
"init_position_mode": "antithetic",
"position_radius": 0.5,
"initial_velocity_radius": 0.5,
"mutation_prob": 0.02,
"reset_velocity_radius": 0.02,
"reflective_bound": 6.0,
"projection_seed": null,
"latent_dim": "full",
"description": "test sufficient bound expansion"
},
"G6": {
"config_id": "G6",
"scale_type": "per_tensor_sd",
"init_position_mode": "antithetic",
"position_radius": 1.5,
"initial_velocity_radius": 0.5,
"mutation_prob": 0.02,
"reset_velocity_radius": 0.02,
"reflective_bound": 6.0,
"projection_seed": null,
"latent_dim": "full",
"description": "test broader normalized initialization"
},
"G8": {
"config_id": "G8",
"scale_type": "optimizer_default",
"init_position_mode": "independent",
"position_radius": 0.05,
"initial_velocity_radius": 0.05,
"mutation_prob": 0.02,
"reset_velocity_radius": 0.02,
"reflective_bound": 3.0,
"projection_seed": null,
"latent_dim": "full",
"description": "retained semantic control (public Optimizer)"
}
},
"workloads": {
"mnist_compact": {
"workload_id": "mnist_compact",
"dataset_name": "mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"model_fingerprint": "d0eee0ffd33088ed",
"data_fingerprint": "26bd2ed5f27e7d24",
"split_fingerprint": "51b289d9f503a9f3",
"description": "MNIST dataset with CompactCNN (9,098 params control)"
},
"mnist_wide": {
"workload_id": "mnist_wide",
"dataset_name": "mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"model_fingerprint": "d474d0ab7cc9f83d",
"data_fingerprint": "26bd2ed5f27e7d24",
"split_fingerprint": "51b289d9f503a9f3",
"description": "MNIST dataset with WideCNN (~55k params larger model axis)"
},
"fashion_compact": {
"workload_id": "fashion_compact",
"dataset_name": "fashion_mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"model_fingerprint": "d0eee0ffd33088ed",
"data_fingerprint": "804ce8cb7d21089f",
"split_fingerprint": "9b30a85c7ff86e4a",
"description": "FashionMNIST dataset with CompactCNN (harder data axis)"
},
"fashion_wide": {
"workload_id": "fashion_wide",
"dataset_name": "fashion_mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"model_fingerprint": "d474d0ab7cc9f83d",
"data_fingerprint": "804ce8cb7d21089f",
"split_fingerprint": "9b30a85c7ff86e4a",
"description": "FashionMNIST dataset with WideCNN (harder data + larger model axis)"
}
},
"untrained_baselines": {
"mnist_compact": {
"val_nll": 2.325095,
"val_accuracy": 6.65,
"val_brier": 0.904644,
"val_ece": 0.060849,
"val_margin": 0.0109,
"objective_loss": 2.325288,
"objective_accuracy": 7.05
},
"mnist_wide": {
"val_nll": 2.32458,
"val_accuracy": 9.31,
"val_brier": 0.904232,
"val_ece": 0.025719,
"val_margin": 0.006578,
"objective_loss": 2.322588,
"objective_accuracy": 9.85
},
"fashion_compact": {
"val_nll": 2.320302,
"val_accuracy": 8.08,
"val_brier": 0.903272,
"val_ece": 0.037873,
"val_margin": 0.005784,
"objective_loss": 2.317111,
"objective_accuracy": 8.65
},
"fashion_wide": {
"val_nll": 2.311222,
"val_accuracy": 8.77,
"val_brier": 0.901676,
"val_ece": 0.029384,
"val_margin": 0.004878,
"objective_loss": 2.311339,
"objective_accuracy": 9.3
}
},
"screen_results": [
{
"workload_id": "mnist_compact",
"method_id": "G0",
"dataset_name": "mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 2.014459,
"gbest_acc": 34.4,
"gbest_val_loss": 2.019988,
"gbest_val_acc": 33.32,
"val_selected_particle_idx": 1,
"val_selected_loss": 2.019988,
"val_selected_acc": 33.32,
"val_metrics": {
"accuracy": 33.32,
"nll": 2.019988,
"brier": 0.834431,
"ece": 0.139305,
"margin": 0.049729
},
"wall_time_sec": 0.9417,
"optimization_wall_time_sec": 0.8673,
"validation_wall_time_sec": 0.0744,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1106883.43,
"is_finite": true
},
{
"workload_id": "mnist_compact",
"method_id": "G5",
"dataset_name": "mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.971574,
"gbest_acc": 32.35,
"gbest_val_loss": 1.959743,
"gbest_val_acc": 33.95,
"val_selected_particle_idx": 9,
"val_selected_loss": 1.959743,
"val_selected_acc": 33.95,
"val_metrics": {
"accuracy": 33.95,
"nll": 1.959743,
"brier": 0.814657,
"ece": 0.092241,
"margin": 0.081765
},
"wall_time_sec": 0.9124,
"optimization_wall_time_sec": 0.8425,
"validation_wall_time_sec": 0.0699,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1139465.88,
"is_finite": true
},
{
"workload_id": "mnist_compact",
"method_id": "G6",
"dataset_name": "mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.941859,
"gbest_acc": 35.35,
"gbest_val_loss": 1.958177,
"gbest_val_acc": 35.03,
"val_selected_particle_idx": 11,
"val_selected_loss": 1.957307,
"val_selected_acc": 35.11,
"val_metrics": {
"accuracy": 35.11,
"nll": 1.957307,
"brier": 0.80199,
"ece": 0.079196,
"margin": 0.101943
},
"wall_time_sec": 0.7827,
"optimization_wall_time_sec": 0.7159,
"validation_wall_time_sec": 0.0668,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1340969.41,
"is_finite": true
},
{
"workload_id": "mnist_compact",
"method_id": "G8",
"dataset_name": "mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.675063,
"gbest_acc": 43.6,
"gbest_val_loss": null,
"gbest_val_acc": null,
"val_selected_particle_idx": null,
"val_selected_loss": 1.690442,
"val_selected_acc": 43.88,
"val_metrics": {
"accuracy": 43.88,
"nll": 1.690442,
"brier": 0.721828,
"ece": 0.046373,
"margin": 0.20173
},
"wall_time_sec": 1.1597,
"optimization_wall_time_sec": 0.975,
"validation_wall_time_sec": 0.1847,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 984615.38,
"is_finite": true
},
{
"workload_id": "mnist_wide",
"method_id": "G0",
"dataset_name": "mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 2.087951,
"gbest_acc": 32.45,
"gbest_val_loss": 2.088156,
"gbest_val_acc": 33.19,
"val_selected_particle_idx": 1,
"val_selected_loss": 2.088156,
"val_selected_acc": 33.19,
"val_metrics": {
"accuracy": 33.19,
"nll": 2.088156,
"brier": 0.841868,
"ece": 0.149299,
"margin": 0.048573
},
"wall_time_sec": 1.8103,
"optimization_wall_time_sec": 1.6122,
"validation_wall_time_sec": 0.1981,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 595459.62,
"is_finite": true
},
{
"workload_id": "mnist_wide",
"method_id": "G5",
"dataset_name": "mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.924357,
"gbest_acc": 39.5,
"gbest_val_loss": 1.913567,
"gbest_val_acc": 39.87,
"val_selected_particle_idx": 1,
"val_selected_loss": 1.913566,
"val_selected_acc": 39.87,
"val_metrics": {
"accuracy": 39.87,
"nll": 1.913566,
"brier": 0.788297,
"ece": 0.114298,
"margin": 0.113113
},
"wall_time_sec": 1.4089,
"optimization_wall_time_sec": 1.2427,
"validation_wall_time_sec": 0.1662,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 772511.47,
"is_finite": true
},
{
"workload_id": "mnist_wide",
"method_id": "G6",
"dataset_name": "mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 2.014518,
"gbest_acc": 27.8,
"gbest_val_loss": 2.017511,
"gbest_val_acc": 29.46,
"val_selected_particle_idx": 2,
"val_selected_loss": 2.017511,
"val_selected_acc": 29.46,
"val_metrics": {
"accuracy": 29.46,
"nll": 2.017511,
"brier": 0.836293,
"ece": 0.08699,
"margin": 0.050265
},
"wall_time_sec": 1.4093,
"optimization_wall_time_sec": 1.2355,
"validation_wall_time_sec": 0.1738,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 777013.35,
"is_finite": true
},
{
"workload_id": "mnist_wide",
"method_id": "G8",
"dataset_name": "mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 2.062368,
"gbest_acc": 20.3,
"gbest_val_loss": null,
"gbest_val_acc": null,
"val_selected_particle_idx": null,
"val_selected_loss": 2.056341,
"val_selected_acc": 20.43,
"val_metrics": {
"accuracy": 20.43,
"nll": 2.056341,
"brier": 0.850317,
"ece": 0.022169,
"margin": 0.052276
},
"wall_time_sec": 1.3145,
"optimization_wall_time_sec": 1.1228,
"validation_wall_time_sec": 0.1917,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 855005.34,
"is_finite": true
},
{
"workload_id": "fashion_compact",
"method_id": "G0",
"dataset_name": "fashion_mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.96974,
"gbest_acc": 28.15,
"gbest_val_loss": 1.969548,
"gbest_val_acc": 28.25,
"val_selected_particle_idx": 0,
"val_selected_loss": 1.969548,
"val_selected_acc": 28.25,
"val_metrics": {
"accuracy": 28.25,
"nll": 1.969548,
"brier": 0.831504,
"ece": 0.090757,
"margin": 0.055678
},
"wall_time_sec": 0.8633,
"optimization_wall_time_sec": 0.7869,
"validation_wall_time_sec": 0.0764,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1219977.13,
"is_finite": true
},
{
"workload_id": "fashion_compact",
"method_id": "G5",
"dataset_name": "fashion_mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.874629,
"gbest_acc": 36.6,
"gbest_val_loss": 1.893233,
"gbest_val_acc": 35.54,
"val_selected_particle_idx": 3,
"val_selected_loss": 1.892402,
"val_selected_acc": 34.31,
"val_metrics": {
"accuracy": 34.31,
"nll": 1.892402,
"brier": 0.798486,
"ece": 0.105844,
"margin": 0.071749
},
"wall_time_sec": 0.9283,
"optimization_wall_time_sec": 0.8554,
"validation_wall_time_sec": 0.073,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1122281.97,
"is_finite": true
},
{
"workload_id": "fashion_compact",
"method_id": "G6",
"dataset_name": "fashion_mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.844914,
"gbest_acc": 33.5,
"gbest_val_loss": 1.824273,
"gbest_val_acc": 35.2,
"val_selected_particle_idx": 8,
"val_selected_loss": 1.824273,
"val_selected_acc": 35.2,
"val_metrics": {
"accuracy": 35.2,
"nll": 1.824273,
"brier": 0.787638,
"ece": 0.046928,
"margin": 0.169004
},
"wall_time_sec": 0.9054,
"optimization_wall_time_sec": 0.8331,
"validation_wall_time_sec": 0.0723,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1152322.65,
"is_finite": true
},
{
"workload_id": "fashion_compact",
"method_id": "G8",
"dataset_name": "fashion_mnist",
"model_name": "compact_cnn",
"parameter_count": 9098,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.895269,
"gbest_acc": 28.85,
"gbest_val_loss": null,
"gbest_val_acc": null,
"val_selected_particle_idx": null,
"val_selected_loss": 1.88018,
"val_selected_acc": 30.85,
"val_metrics": {
"accuracy": 30.85,
"nll": 1.88018,
"brier": 0.800592,
"ece": 0.043279,
"margin": 0.16498
},
"wall_time_sec": 1.0335,
"optimization_wall_time_sec": 0.8046,
"validation_wall_time_sec": 0.2289,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 1193139.45,
"is_finite": true
},
{
"workload_id": "fashion_wide",
"method_id": "G0",
"dataset_name": "fashion_mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.98436,
"gbest_acc": 36.05,
"gbest_val_loss": 1.981317,
"gbest_val_acc": 36.39,
"val_selected_particle_idx": 8,
"val_selected_loss": 1.981316,
"val_selected_acc": 36.39,
"val_metrics": {
"accuracy": 36.39,
"nll": 1.981316,
"brier": 0.826711,
"ece": 0.180482,
"margin": 0.042564
},
"wall_time_sec": 1.9694,
"optimization_wall_time_sec": 1.7767,
"validation_wall_time_sec": 0.1927,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 540327.57,
"is_finite": true
},
{
"workload_id": "fashion_wide",
"method_id": "G5",
"dataset_name": "fashion_mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.938569,
"gbest_acc": 33.3,
"gbest_val_loss": 1.945862,
"gbest_val_acc": 31.97,
"val_selected_particle_idx": 0,
"val_selected_loss": 1.945862,
"val_selected_acc": 31.97,
"val_metrics": {
"accuracy": 31.97,
"nll": 1.945862,
"brier": 0.818283,
"ece": 0.114404,
"margin": 0.049718
},
"wall_time_sec": 1.335,
"optimization_wall_time_sec": 1.189,
"validation_wall_time_sec": 0.146,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 807401.18,
"is_finite": true
},
{
"workload_id": "fashion_wide",
"method_id": "G6",
"dataset_name": "fashion_mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.960139,
"gbest_acc": 28.1,
"gbest_val_loss": 1.975825,
"gbest_val_acc": 28.64,
"val_selected_particle_idx": 10,
"val_selected_loss": 1.975825,
"val_selected_acc": 28.64,
"val_metrics": {
"accuracy": 28.64,
"nll": 1.975825,
"brier": 0.825556,
"ece": 0.038186,
"margin": 0.084988
},
"wall_time_sec": 1.4862,
"optimization_wall_time_sec": 1.3055,
"validation_wall_time_sec": 0.1807,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 17,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 735350.44,
"is_finite": true
},
{
"workload_id": "fashion_wide",
"method_id": "G8",
"dataset_name": "fashion_mnist",
"model_name": "wide_cnn",
"parameter_count": 55338,
"subset_size": 2000,
"particles": 12,
"epochs": 40,
"seed": 91,
"gbest_loss": 1.976977,
"gbest_acc": 28.45,
"gbest_val_loss": null,
"gbest_val_acc": null,
"val_selected_particle_idx": null,
"val_selected_loss": 1.987923,
"val_selected_acc": 28.69,
"val_metrics": {
"accuracy": 28.69,
"nll": 1.987923,
"brier": 0.829408,
"ece": 0.053144,
"margin": 0.069324
},
"wall_time_sec": 1.2958,
"optimization_wall_time_sec": 1.1088,
"validation_wall_time_sec": 0.187,
"total_queries": 480,
"total_sample_evaluations": 960000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 865800.87,
"is_finite": true
}
],
"normalized_method_selections": {
"mnist_compact": "G6",
"mnist_wide": "G5",
"fashion_compact": "G6",
"fashion_wide": "G5"
},
"confirmation_results": {
"mnist_compact": {
"G8": {
"workload_id": "mnist_compact",
"method_id": "G8",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 49.153333,
"std": 1.320656,
"median": 48.96,
"iqr": 1.31,
"ci95_t": 3.280728
},
"val_nll": {
"mean": 1.518089,
"std": 0.061523,
"median": 1.512633,
"iqr": 0.061342,
"ci95_t": 0.152834
},
"val_brier": {
"mean": 0.663921,
"std": 0.024244,
"median": 0.659644,
"iqr": 0.02396,
"ci95_t": 0.060226
},
"val_ece": {
"mean": 0.066236,
"std": 0.050095,
"median": 0.060988,
"iqr": 0.049888,
"ci95_t": 0.124443
},
"gbest_loss": {
"mean": 1.518602,
"std": 0.062467,
"median": 1.509299,
"iqr": 0.061945,
"ci95_t": 0.155179
},
"gbest_acc": {
"mean": 48.96,
"std": 1.407551,
"median": 49.28,
"iqr": 1.38,
"ci95_t": 3.496589
},
"wall_time_sec": {
"mean": 4.530167,
"std": 0.073436,
"median": 4.5186,
"iqr": 0.07275,
"ci95_t": 0.182428
},
"optimization_wall_time_sec": {
"mean": 4.3879,
"std": 0.064886,
"median": 4.3684,
"iqr": 0.06265,
"ci95_t": 0.161188
},
"throughput_samples_per_sec": {
"mean": 2188151.803333,
"std": 32164.042335,
"median": 2197600.95,
"iqr": 31105.64,
"ci95_t": 79900.788332
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.459476,
"val_selected_acc": 50.56,
"val_metrics": {
"accuracy": 50.56,
"nll": 1.459476,
"brier": 0.6421,
"ece": 0.018972,
"margin": 0.29628
},
"gbest_loss": 1.461308,
"gbest_acc": 50.18,
"wall_time_sec": 4.5186,
"optimization_wall_time_sec": 4.3684,
"validation_wall_time_sec": 0.1502,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 2197600.95
},
{
"seed": 102,
"val_selected_loss": 1.582159,
"val_selected_acc": 47.94,
"val_metrics": {
"accuracy": 47.94,
"nll": 1.582159,
"brier": 0.690019,
"ece": 0.118748,
"margin": 0.178296
},
"gbest_loss": 1.585199,
"gbest_acc": 47.42,
"wall_time_sec": 4.4632,
"optimization_wall_time_sec": 4.335,
"validation_wall_time_sec": 0.1282,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 2214532.87
},
{
"seed": 103,
"val_selected_loss": 1.512633,
"val_selected_acc": 48.96,
"val_metrics": {
"accuracy": 48.96,
"nll": 1.512633,
"brier": 0.659644,
"ece": 0.060988,
"margin": 0.242224
},
"gbest_loss": 1.509299,
"gbest_acc": 49.28,
"wall_time_sec": 4.6087,
"optimization_wall_time_sec": 4.4603,
"validation_wall_time_sec": 0.1484,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 2152321.59
}
]
},
"G6": {
"workload_id": "mnist_compact",
"method_id": "G6",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 45.966667,
"std": 3.545438,
"median": 46.62,
"iqr": 3.5,
"ci95_t": 8.807454
},
"val_nll": {
"mean": 1.642081,
"std": 0.091398,
"median": 1.610062,
"iqr": 0.08709,
"ci95_t": 0.227048
},
"val_brier": {
"mean": 0.704608,
"std": 0.0315,
"median": 0.68996,
"iqr": 0.028833,
"ci95_t": 0.07825
},
"val_ece": {
"mean": 0.115422,
"std": 0.039961,
"median": 0.116838,
"iqr": 0.039942,
"ci95_t": 0.09927
},
"gbest_loss": {
"mean": 1.641447,
"std": 0.089036,
"median": 1.60673,
"iqr": 0.083806,
"ci95_t": 0.22118
},
"gbest_acc": {
"mean": 46.056667,
"std": 3.536444,
"median": 46.45,
"iqr": 3.52,
"ci95_t": 8.785109
},
"wall_time_sec": {
"mean": 5.2706,
"std": 1.015911,
"median": 4.8932,
"iqr": 0.9619,
"ci95_t": 2.52369
},
"optimization_wall_time_sec": {
"mean": 5.159533,
"std": 0.984816,
"median": 4.7957,
"iqr": 0.93305,
"ci95_t": 2.446446
},
"throughput_samples_per_sec": {
"mean": 1903152.163333,
"std": 334907.322507,
"median": 2001793.27,
"iqr": 323829.22,
"ci95_t": 831965.049903
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.745181,
"val_selected_acc": 42.14,
"val_metrics": {
"accuracy": 42.14,
"nll": 1.745181,
"brier": 0.740764,
"ece": 0.116838,
"margin": 0.131853
},
"gbest_loss": 1.742612,
"gbest_acc": 42.34,
"wall_time_sec": 4.8932,
"optimization_wall_time_sec": 4.7957,
"validation_wall_time_sec": 0.0975,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 2001793.27
},
{
"seed": 102,
"val_selected_loss": 1.610062,
"val_selected_acc": 49.14,
"val_metrics": {
"accuracy": 49.14,
"nll": 1.610062,
"brier": 0.68996,
"ece": 0.154657,
"margin": 0.160339
},
"gbest_loss": 1.60673,
"gbest_acc": 49.38,
"wall_time_sec": 6.4212,
"optimization_wall_time_sec": 6.2745,
"validation_wall_time_sec": 0.1467,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 1530002.39
},
{
"seed": 103,
"val_selected_loss": 1.571001,
"val_selected_acc": 46.62,
"val_metrics": {
"accuracy": 46.62,
"nll": 1.571001,
"brier": 0.683099,
"ece": 0.074772,
"margin": 0.208291
},
"gbest_loss": 1.575,
"gbest_acc": 46.45,
"wall_time_sec": 4.4974,
"optimization_wall_time_sec": 4.4084,
"validation_wall_time_sec": 0.089,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 2177660.83
}
]
}
},
"mnist_wide": {
"G8": {
"workload_id": "mnist_wide",
"method_id": "G8",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 41.03,
"std": 4.055083,
"median": 41.36,
"iqr": 4.045,
"ci95_t": 10.073496
},
"val_nll": {
"mean": 1.733351,
"std": 0.084331,
"median": 1.733455,
"iqr": 0.08433,
"ci95_t": 0.209491
},
"val_brier": {
"mean": 0.739941,
"std": 0.029512,
"median": 0.734463,
"iqr": 0.029129,
"ci95_t": 0.073314
},
"val_ece": {
"mean": 0.088906,
"std": 0.015336,
"median": 0.083434,
"iqr": 0.014586,
"ci95_t": 0.038097
},
"gbest_loss": {
"mean": 1.732307,
"std": 0.09299,
"median": 1.720824,
"iqr": 0.092456,
"ci95_t": 0.231002
},
"gbest_acc": {
"mean": 40.866667,
"std": 3.390197,
"median": 40.44,
"iqr": 3.37,
"ci95_t": 8.421808
},
"wall_time_sec": {
"mean": 9.4569,
"std": 0.542029,
"median": 9.5612,
"iqr": 0.53445,
"ci95_t": 1.34649
},
"optimization_wall_time_sec": {
"mean": 9.2436,
"std": 0.534734,
"median": 9.3462,
"iqr": 0.5273,
"ci95_t": 1.328367
},
"throughput_samples_per_sec": {
"mean": 1040918.586667,
"std": 61275.63655,
"median": 1027155.42,
"iqr": 60105.2,
"ci95_t": 152218.791868
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.733455,
"val_selected_acc": 41.36,
"val_metrics": {
"accuracy": 41.36,
"nll": 1.733455,
"brier": 0.734463,
"ece": 0.077057,
"margin": 0.152205
},
"gbest_loss": 1.720824,
"gbest_acc": 40.44,
"wall_time_sec": 8.8703,
"optimization_wall_time_sec": 8.665,
"validation_wall_time_sec": 0.2053,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 1107905.37
},
{
"seed": 102,
"val_selected_loss": 1.648968,
"val_selected_acc": 44.91,
"val_metrics": {
"accuracy": 44.91,
"nll": 1.648968,
"brier": 0.713552,
"ece": 0.106228,
"margin": 0.14736
},
"gbest_loss": 1.645592,
"gbest_acc": 44.45,
"wall_time_sec": 9.5612,
"optimization_wall_time_sec": 9.3462,
"validation_wall_time_sec": 0.215,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 1027155.42
},
{
"seed": 103,
"val_selected_loss": 1.817629,
"val_selected_acc": 36.82,
"val_metrics": {
"accuracy": 36.82,
"nll": 1.817629,
"brier": 0.771809,
"ece": 0.083434,
"margin": 0.112362
},
"gbest_loss": 1.830505,
"gbest_acc": 37.71,
"wall_time_sec": 9.9392,
"optimization_wall_time_sec": 9.7196,
"validation_wall_time_sec": 0.2196,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 987694.97
}
]
},
"G5": {
"workload_id": "mnist_wide",
"method_id": "G5",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 43.86,
"std": 3.512222,
"median": 43.1,
"iqr": 3.45,
"ci95_t": 8.724938
},
"val_nll": {
"mean": 1.721259,
"std": 0.098656,
"median": 1.66647,
"iqr": 0.086496,
"ci95_t": 0.245077
},
"val_brier": {
"mean": 0.734178,
"std": 0.034098,
"median": 0.737217,
"iqr": 0.033996,
"ci95_t": 0.084704
},
"val_ece": {
"mean": 0.150847,
"std": 0.013927,
"median": 0.144481,
"iqr": 0.012789,
"ci95_t": 0.034596
},
"gbest_loss": {
"mean": 1.716147,
"std": 0.098467,
"median": 1.667192,
"iqr": 0.088872,
"ci95_t": 0.244608
},
"gbest_acc": {
"mean": 44.03,
"std": 4.013776,
"median": 43.43,
"iqr": 3.98,
"ci95_t": 9.970883
},
"wall_time_sec": {
"mean": 9.868467,
"std": 0.070779,
"median": 9.8397,
"iqr": 0.06625,
"ci95_t": 0.175827
},
"optimization_wall_time_sec": {
"mean": 9.6602,
"std": 0.069128,
"median": 9.6324,
"iqr": 0.0648,
"ci95_t": 0.171725
},
"throughput_samples_per_sec": {
"mean": 993802.05,
"std": 7086.000756,
"median": 996636.35,
"iqr": 6647.29,
"ci95_t": 17602.795091
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.66647,
"val_selected_acc": 43.1,
"val_metrics": {
"accuracy": 43.1,
"nll": 1.66647,
"brier": 0.737217,
"ece": 0.144481,
"margin": 0.102556
},
"gbest_loss": 1.667192,
"gbest_acc": 43.43,
"wall_time_sec": 9.9491,
"optimization_wall_time_sec": 9.7389,
"validation_wall_time_sec": 0.2102,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 985737.61
},
{
"seed": 102,
"val_selected_loss": 1.662158,
"val_selected_acc": 47.69,
"val_metrics": {
"accuracy": 47.69,
"nll": 1.662158,
"brier": 0.698663,
"ece": 0.166819,
"margin": 0.156783
},
"gbest_loss": 1.651753,
"gbest_acc": 48.31,
"wall_time_sec": 9.8166,
"optimization_wall_time_sec": 9.6093,
"validation_wall_time_sec": 0.2073,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 999032.19
},
{
"seed": 103,
"val_selected_loss": 1.83515,
"val_selected_acc": 40.79,
"val_metrics": {
"accuracy": 40.79,
"nll": 1.83515,
"brier": 0.766655,
"ece": 0.141241,
"margin": 0.108132
},
"gbest_loss": 1.829497,
"gbest_acc": 40.35,
"wall_time_sec": 9.8397,
"optimization_wall_time_sec": 9.6324,
"validation_wall_time_sec": 0.2073,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 996636.35
}
]
}
},
"fashion_compact": {
"G8": {
"workload_id": "fashion_compact",
"method_id": "G8",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 47.003333,
"std": 5.2259,
"median": 44.5,
"iqr": 4.755,
"ci95_t": 12.981998
},
"val_nll": {
"mean": 1.511217,
"std": 0.168628,
"median": 1.593665,
"iqr": 0.152765,
"ci95_t": 0.4189
},
"val_brier": {
"mean": 0.670829,
"std": 0.047121,
"median": 0.69484,
"iqr": 0.042284,
"ci95_t": 0.117055
},
"val_ece": {
"mean": 0.051117,
"std": 0.028629,
"median": 0.042956,
"iqr": 0.027743,
"ci95_t": 0.071118
},
"gbest_loss": {
"mean": 1.505129,
"std": 0.173608,
"median": 1.585615,
"iqr": 0.159001,
"ci95_t": 0.431271
},
"gbest_acc": {
"mean": 47.446667,
"std": 5.162212,
"median": 44.87,
"iqr": 4.655,
"ci95_t": 12.823787
},
"wall_time_sec": {
"mean": 5.0146,
"std": 0.07329,
"median": 4.9742,
"iqr": 0.0644,
"ci95_t": 0.182065
},
"optimization_wall_time_sec": {
"mean": 4.860267,
"std": 0.074059,
"median": 4.8208,
"iqr": 0.0657,
"ci95_t": 0.183974
},
"throughput_samples_per_sec": {
"mean": 1975503.406667,
"std": 29841.726913,
"median": 1991370.73,
"iqr": 26489.615,
"ci95_t": 74131.773634
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.317228,
"val_selected_acc": 53.01,
"val_metrics": {
"accuracy": 53.01,
"nll": 1.317228,
"brier": 0.616539,
"ece": 0.042956,
"margin": 0.295591
},
"gbest_loss": 1.305885,
"gbest_acc": 53.39,
"wall_time_sec": 4.9704,
"optimization_wall_time_sec": 4.8143,
"validation_wall_time_sec": 0.1561,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 1994059.36
},
{
"seed": 102,
"val_selected_loss": 1.593665,
"val_selected_acc": 43.5,
"val_metrics": {
"accuracy": 43.5,
"nll": 1.593665,
"brier": 0.69484,
"ece": 0.027455,
"margin": 0.271024
},
"gbest_loss": 1.585615,
"gbest_acc": 44.08,
"wall_time_sec": 5.0992,
"optimization_wall_time_sec": 4.9457,
"validation_wall_time_sec": 0.1534,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 1941080.13
},
{
"seed": 103,
"val_selected_loss": 1.622758,
"val_selected_acc": 44.5,
"val_metrics": {
"accuracy": 44.5,
"nll": 1.622758,
"brier": 0.701107,
"ece": 0.08294,
"margin": 0.177938
},
"gbest_loss": 1.623887,
"gbest_acc": 44.87,
"wall_time_sec": 4.9742,
"optimization_wall_time_sec": 4.8208,
"validation_wall_time_sec": 0.1534,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2219912,
"throughput_samples_per_sec": 1991370.73
}
]
},
"G6": {
"workload_id": "fashion_compact",
"method_id": "G6",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 45.616667,
"std": 3.585392,
"median": 44.86,
"iqr": 3.525,
"ci95_t": 8.906704
},
"val_nll": {
"mean": 1.584443,
"std": 0.053077,
"median": 1.593562,
"iqr": 0.052487,
"ci95_t": 0.131853
},
"val_brier": {
"mean": 0.699684,
"std": 0.023018,
"median": 0.699068,
"iqr": 0.023011,
"ci95_t": 0.05718
},
"val_ece": {
"mean": 0.098265,
"std": 0.040761,
"median": 0.075459,
"iqr": 0.035656,
"ci95_t": 0.101256
},
"gbest_loss": {
"mean": 1.577682,
"std": 0.060531,
"median": 1.589026,
"iqr": 0.059728,
"ci95_t": 0.150369
},
"gbest_acc": {
"mean": 45.83,
"std": 3.77342,
"median": 44.62,
"iqr": 3.625,
"ci95_t": 9.373798
},
"wall_time_sec": {
"mean": 4.839267,
"std": 0.025908,
"median": 4.8501,
"iqr": 0.02415,
"ci95_t": 0.064361
},
"optimization_wall_time_sec": {
"mean": 4.7373,
"std": 0.0259,
"median": 4.7484,
"iqr": 0.02405,
"ci95_t": 0.06434
},
"throughput_samples_per_sec": {
"mean": 2026511.273333,
"std": 11111.347421,
"median": 2021733.64,
"iqr": 10312.26,
"ci95_t": 27602.420402
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.593562,
"val_selected_acc": 44.86,
"val_metrics": {
"accuracy": 44.86,
"nll": 1.593562,
"brier": 0.699068,
"ece": 0.075459,
"margin": 0.179908
},
"gbest_loss": 1.589026,
"gbest_acc": 44.62,
"wall_time_sec": 4.8501,
"optimization_wall_time_sec": 4.7484,
"validation_wall_time_sec": 0.1017,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 2021733.64
},
{
"seed": 102,
"val_selected_loss": 1.63237,
"val_selected_acc": 42.47,
"val_metrics": {
"accuracy": 42.47,
"nll": 1.63237,
"brier": 0.723003,
"ece": 0.074012,
"margin": 0.176538
},
"gbest_loss": 1.631738,
"gbest_acc": 42.81,
"wall_time_sec": 4.8097,
"optimization_wall_time_sec": 4.7077,
"validation_wall_time_sec": 0.102,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 2039212.35
},
{
"seed": 103,
"val_selected_loss": 1.527397,
"val_selected_acc": 49.52,
"val_metrics": {
"accuracy": 49.52,
"nll": 1.527397,
"brier": 0.67698,
"ece": 0.145324,
"margin": 0.156547
},
"gbest_loss": 1.512281,
"gbest_acc": 50.06,
"wall_time_sec": 4.858,
"optimization_wall_time_sec": 4.7558,
"validation_wall_time_sec": 0.1022,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 2183520,
"throughput_samples_per_sec": 2018587.83
}
]
}
},
"fashion_wide": {
"G8": {
"workload_id": "fashion_wide",
"method_id": "G8",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 41.666667,
"std": 2.269016,
"median": 40.94,
"iqr": 2.18,
"ci95_t": 5.63661
},
"val_nll": {
"mean": 1.615937,
"std": 0.036692,
"median": 1.617185,
"iqr": 0.036676,
"ci95_t": 0.091149
},
"val_brier": {
"mean": 0.716583,
"std": 0.015392,
"median": 0.719193,
"iqr": 0.015226,
"ci95_t": 0.038237
},
"val_ece": {
"mean": 0.039104,
"std": 0.003431,
"median": 0.039736,
"iqr": 0.003387,
"ci95_t": 0.008523
},
"gbest_loss": {
"mean": 1.612987,
"std": 0.031992,
"median": 1.612069,
"iqr": 0.031983,
"ci95_t": 0.079474
},
"gbest_acc": {
"mean": 41.346667,
"std": 2.309899,
"median": 40.59,
"iqr": 2.215,
"ci95_t": 5.73817
},
"wall_time_sec": {
"mean": 9.499333,
"std": 0.068809,
"median": 9.5038,
"iqr": 0.0687,
"ci95_t": 0.170932
},
"optimization_wall_time_sec": {
"mean": 9.280267,
"std": 0.067396,
"median": 9.2862,
"iqr": 0.0672,
"ci95_t": 0.167423
},
"throughput_samples_per_sec": {
"mean": 1034489.443333,
"std": 7520.124937,
"median": 1033792.08,
"iqr": 7495.835,
"ci95_t": 18681.231187
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.617185,
"val_selected_acc": 40.94,
"val_metrics": {
"accuracy": 40.94,
"nll": 1.617185,
"brier": 0.719193,
"ece": 0.039736,
"margin": 0.182429
},
"gbest_loss": 1.612069,
"gbest_acc": 40.59,
"wall_time_sec": 9.4284,
"optimization_wall_time_sec": 9.2101,
"validation_wall_time_sec": 0.2183,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 1042333.96
},
{
"seed": 102,
"val_selected_loss": 1.651989,
"val_selected_acc": 39.85,
"val_metrics": {
"accuracy": 39.85,
"nll": 1.651989,
"brier": 0.730504,
"ece": 0.035401,
"margin": 0.165291
},
"gbest_loss": 1.645429,
"gbest_acc": 39.51,
"wall_time_sec": 9.5038,
"optimization_wall_time_sec": 9.2862,
"validation_wall_time_sec": 0.2176,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 1033792.08
},
{
"seed": 103,
"val_selected_loss": 1.578637,
"val_selected_acc": 44.21,
"val_metrics": {
"accuracy": 44.21,
"nll": 1.578637,
"brier": 0.700053,
"ece": 0.042175,
"margin": 0.190905
},
"gbest_loss": 1.581464,
"gbest_acc": 43.94,
"wall_time_sec": 9.5658,
"optimization_wall_time_sec": 9.3445,
"validation_wall_time_sec": 0.2213,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 13,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13502472,
"throughput_samples_per_sec": 1027342.29
}
]
},
"G5": {
"workload_id": "fashion_wide",
"method_id": "G5",
"subset_size": 10000,
"particles": 12,
"epochs": 80,
"seeds": [
101,
102,
103
],
"stats": {
"val_acc": {
"mean": 46.31,
"std": 7.568494,
"median": 48.76,
"iqr": 7.265,
"ci95_t": 18.801388
},
"val_nll": {
"mean": 1.525747,
"std": 0.149706,
"median": 1.459695,
"iqr": 0.138346,
"ci95_t": 0.371893
},
"val_brier": {
"mean": 0.695206,
"std": 0.057206,
"median": 0.675207,
"iqr": 0.054522,
"ci95_t": 0.14211
},
"val_ece": {
"mean": 0.135458,
"std": 0.015271,
"median": 0.132617,
"iqr": 0.015072,
"ci95_t": 0.037936
},
"gbest_loss": {
"mean": 1.524557,
"std": 0.150724,
"median": 1.462688,
"iqr": 0.14088,
"ci95_t": 0.374424
},
"gbest_acc": {
"mean": 46.256667,
"std": 8.185019,
"median": 49.01,
"iqr": 7.83,
"ci95_t": 20.332937
},
"wall_time_sec": {
"mean": 10.038833,
"std": 0.159562,
"median": 10.0574,
"iqr": 0.15875,
"ci95_t": 0.396379
},
"optimization_wall_time_sec": {
"mean": 9.8224,
"std": 0.154476,
"median": 9.8419,
"iqr": 0.15355,
"ci95_t": 0.383744
},
"throughput_samples_per_sec": {
"mean": 977519.543333,
"std": 15420.239061,
"median": 975421.41,
"iqr": 15312.81,
"ci95_t": 38306.418218
}
},
"per_seed_runs": [
{
"seed": 101,
"val_selected_loss": 1.420427,
"val_selected_acc": 52.35,
"val_metrics": {
"accuracy": 52.35,
"nll": 1.420427,
"brier": 0.650684,
"ece": 0.15195,
"margin": 0.168393
},
"gbest_loss": 1.414612,
"gbest_acc": 52.71,
"wall_time_sec": 9.8708,
"optimization_wall_time_sec": 9.6591,
"validation_wall_time_sec": 0.2117,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 993881.42
},
{
"seed": 102,
"val_selected_loss": 1.459695,
"val_selected_acc": 48.76,
"val_metrics": {
"accuracy": 48.76,
"nll": 1.459695,
"brier": 0.675207,
"ece": 0.132617,
"margin": 0.141615
},
"gbest_loss": 1.462688,
"gbest_acc": 49.01,
"wall_time_sec": 10.0574,
"optimization_wall_time_sec": 9.8419,
"validation_wall_time_sec": 0.2155,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 975421.41
},
{
"seed": 103,
"val_selected_loss": 1.697119,
"val_selected_acc": 37.82,
"val_metrics": {
"accuracy": 37.82,
"nll": 1.697119,
"brier": 0.759727,
"ece": 0.121807,
"margin": 0.068249
},
"gbest_loss": 1.696371,
"gbest_acc": 37.05,
"wall_time_sec": 10.1883,
"optimization_wall_time_sec": 9.9662,
"validation_wall_time_sec": 0.2221,
"total_queries": 960,
"total_sample_evaluations": 9600000,
"validation_evaluations": 21,
"official_test_evaluations": 0,
"core_swarm_state_bytes": 13281120,
"throughput_samples_per_sec": 963255.8
}
]
}
}
},
"feasibility_evaluations": {
"mnist_compact": {
"G8": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.325095,
"target_val_nll_threshold": 1.860076,
"baseline_val_acc": 6.65,
"target_val_acc_threshold": 26.65,
"mean_val_nll": 1.518089,
"mean_val_acc": 49.1533
},
"G6": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.325095,
"target_val_nll_threshold": 1.860076,
"baseline_val_acc": 6.65,
"target_val_acc_threshold": 26.65,
"mean_val_nll": 1.642081,
"mean_val_acc": 45.9667
}
},
"mnist_wide": {
"G8": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.32458,
"target_val_nll_threshold": 1.859664,
"baseline_val_acc": 9.31,
"target_val_acc_threshold": 29.31,
"mean_val_nll": 1.733351,
"mean_val_acc": 41.03
},
"G5": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.32458,
"target_val_nll_threshold": 1.859664,
"baseline_val_acc": 9.31,
"target_val_acc_threshold": 29.31,
"mean_val_nll": 1.721259,
"mean_val_acc": 43.86
}
},
"fashion_compact": {
"G8": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.320302,
"target_val_nll_threshold": 1.856242,
"baseline_val_acc": 8.08,
"target_val_acc_threshold": 28.08,
"mean_val_nll": 1.511217,
"mean_val_acc": 47.0033
},
"G6": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.320302,
"target_val_nll_threshold": 1.856242,
"baseline_val_acc": 8.08,
"target_val_acc_threshold": 28.08,
"mean_val_nll": 1.584443,
"mean_val_acc": 45.6167
}
},
"fashion_wide": {
"G8": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.311222,
"target_val_nll_threshold": 1.848978,
"baseline_val_acc": 8.77,
"target_val_acc_threshold": 28.77,
"mean_val_nll": 1.615937,
"mean_val_acc": 41.6667
},
"G5": {
"execution_feasible": true,
"optimization_feasible": true,
"baseline_val_nll": 2.311222,
"target_val_nll_threshold": 1.848978,
"baseline_val_acc": 8.77,
"target_val_acc_threshold": 28.77,
"mean_val_nll": 1.525747,
"mean_val_acc": 46.31
}
}
},
"resource_totals": {
"total_queries": 30720,
"total_sample_evaluations": 245760000,
"summed_optimization_wall_time_sec": 188.7283,
"summed_validation_wall_time_sec": 6.3826,
"summed_recorded_run_wall_time_sec": 195.1109,
"elapsed_current_process_wall_time_sec": 176.3431
}
}
Reference in New Issue View Git Blame Copy Permalink