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

feat: modernize PSO and add convergence research

Browse Source 
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
This commit is contained in:
jung-geun
2026-09-07 22:03:25 +09:00
parent 9745bb7ad4
commit 813433000a
120 changed files with 198148 additions and 4745 deletions
Download Patch File Download Diff File Expand all files Collapse all files
benchmark_results/pso_v6_heavy_tasks.csv
+164
View File
@@ -0,0 +1,164 @@
section,workload,method,metric,value
protocol,global,all,version,HEAVY-TASK-PSO-V6 1.0.0
protocol,global,all,official_test_data_loaded,False
protocol,global,all,official_test_evaluations,0
baseline,mnist_compact,untrained,val_nll,2.325095
baseline,mnist_compact,untrained,val_accuracy,6.65
baseline,mnist_wide,untrained,val_nll,2.32458
baseline,mnist_wide,untrained,val_accuracy,9.31
baseline,fashion_compact,untrained,val_nll,2.320302
baseline,fashion_compact,untrained,val_accuracy,8.08
baseline,fashion_wide,untrained,val_nll,2.311222
baseline,fashion_wide,untrained,val_accuracy,8.77
screen,mnist_compact,G0,val_nll,2.019988
screen,mnist_compact,G0,val_acc,33.32
screen,mnist_compact,G0,gbest_loss,2.014459
screen,mnist_compact,G0,gbest_acc,34.4
screen,mnist_compact,G0,throughput_sps,1106883.43
screen,mnist_compact,G5,val_nll,1.959743
screen,mnist_compact,G5,val_acc,33.95
screen,mnist_compact,G5,gbest_loss,1.971574
screen,mnist_compact,G5,gbest_acc,32.35
screen,mnist_compact,G5,throughput_sps,1139465.88
screen,mnist_compact,G6,val_nll,1.957307
screen,mnist_compact,G6,val_acc,35.11
screen,mnist_compact,G6,gbest_loss,1.941859
screen,mnist_compact,G6,gbest_acc,35.35
screen,mnist_compact,G6,throughput_sps,1340969.41
screen,mnist_compact,G8,val_nll,1.690442
screen,mnist_compact,G8,val_acc,43.88
screen,mnist_compact,G8,gbest_loss,1.675063
screen,mnist_compact,G8,gbest_acc,43.6
screen,mnist_compact,G8,throughput_sps,984615.38
screen,mnist_wide,G0,val_nll,2.088156
screen,mnist_wide,G0,val_acc,33.19
screen,mnist_wide,G0,gbest_loss,2.087951
screen,mnist_wide,G0,gbest_acc,32.45
screen,mnist_wide,G0,throughput_sps,595459.62
screen,mnist_wide,G5,val_nll,1.913566
screen,mnist_wide,G5,val_acc,39.87
screen,mnist_wide,G5,gbest_loss,1.924357
screen,mnist_wide,G5,gbest_acc,39.5
screen,mnist_wide,G5,throughput_sps,772511.47
screen,mnist_wide,G6,val_nll,2.017511
screen,mnist_wide,G6,val_acc,29.46
screen,mnist_wide,G6,gbest_loss,2.014518
screen,mnist_wide,G6,gbest_acc,27.8
screen,mnist_wide,G6,throughput_sps,777013.35
screen,mnist_wide,G8,val_nll,2.056341
screen,mnist_wide,G8,val_acc,20.43
screen,mnist_wide,G8,gbest_loss,2.062368
screen,mnist_wide,G8,gbest_acc,20.3
screen,mnist_wide,G8,throughput_sps,855005.34
screen,fashion_compact,G0,val_nll,1.969548
screen,fashion_compact,G0,val_acc,28.25
screen,fashion_compact,G0,gbest_loss,1.96974
screen,fashion_compact,G0,gbest_acc,28.15
screen,fashion_compact,G0,throughput_sps,1219977.13
screen,fashion_compact,G5,val_nll,1.892402
screen,fashion_compact,G5,val_acc,34.31
screen,fashion_compact,G5,gbest_loss,1.874629
screen,fashion_compact,G5,gbest_acc,36.6
screen,fashion_compact,G5,throughput_sps,1122281.97
screen,fashion_compact,G6,val_nll,1.824273
screen,fashion_compact,G6,val_acc,35.2
screen,fashion_compact,G6,gbest_loss,1.844914
screen,fashion_compact,G6,gbest_acc,33.5
screen,fashion_compact,G6,throughput_sps,1152322.65
screen,fashion_compact,G8,val_nll,1.88018
screen,fashion_compact,G8,val_acc,30.85
screen,fashion_compact,G8,gbest_loss,1.895269
screen,fashion_compact,G8,gbest_acc,28.85
screen,fashion_compact,G8,throughput_sps,1193139.45
screen,fashion_wide,G0,val_nll,1.981316
screen,fashion_wide,G0,val_acc,36.39
screen,fashion_wide,G0,gbest_loss,1.98436
screen,fashion_wide,G0,gbest_acc,36.05
screen,fashion_wide,G0,throughput_sps,540327.57
screen,fashion_wide,G5,val_nll,1.945862
screen,fashion_wide,G5,val_acc,31.97
screen,fashion_wide,G5,gbest_loss,1.938569
screen,fashion_wide,G5,gbest_acc,33.3
screen,fashion_wide,G5,throughput_sps,807401.18
screen,fashion_wide,G6,val_nll,1.975825
screen,fashion_wide,G6,val_acc,28.64
screen,fashion_wide,G6,gbest_loss,1.960139
screen,fashion_wide,G6,gbest_acc,28.1
screen,fashion_wide,G6,throughput_sps,735350.44
screen,fashion_wide,G8,val_nll,1.987923
screen,fashion_wide,G8,val_acc,28.69
screen,fashion_wide,G8,gbest_loss,1.976977
screen,fashion_wide,G8,gbest_acc,28.45
screen,fashion_wide,G8,throughput_sps,865800.87
confirm,mnist_compact,G8,val_acc_mean,49.153333
confirm,mnist_compact,G8,val_acc_std,1.320656
confirm,mnist_compact,G8,val_nll_mean,1.518089
confirm,mnist_compact,G8,val_nll_std,0.061523
confirm,mnist_compact,G8,wall_time_sec_mean,4.530167
confirm,mnist_compact,G6,val_acc_mean,45.966667
confirm,mnist_compact,G6,val_acc_std,3.545438
confirm,mnist_compact,G6,val_nll_mean,1.642081
confirm,mnist_compact,G6,val_nll_std,0.091398
confirm,mnist_compact,G6,wall_time_sec_mean,5.2706
confirm,mnist_wide,G8,val_acc_mean,41.03
confirm,mnist_wide,G8,val_acc_std,4.055083
confirm,mnist_wide,G8,val_nll_mean,1.733351
confirm,mnist_wide,G8,val_nll_std,0.084331
confirm,mnist_wide,G8,wall_time_sec_mean,9.4569
confirm,mnist_wide,G5,val_acc_mean,43.86
confirm,mnist_wide,G5,val_acc_std,3.512222
confirm,mnist_wide,G5,val_nll_mean,1.721259
confirm,mnist_wide,G5,val_nll_std,0.098656
confirm,mnist_wide,G5,wall_time_sec_mean,9.868467
confirm,fashion_compact,G8,val_acc_mean,47.003333
confirm,fashion_compact,G8,val_acc_std,5.2259
confirm,fashion_compact,G8,val_nll_mean,1.511217
confirm,fashion_compact,G8,val_nll_std,0.168628
confirm,fashion_compact,G8,wall_time_sec_mean,5.0146
confirm,fashion_compact,G6,val_acc_mean,45.616667
confirm,fashion_compact,G6,val_acc_std,3.585392
confirm,fashion_compact,G6,val_nll_mean,1.584443
confirm,fashion_compact,G6,val_nll_std,0.053077
confirm,fashion_compact,G6,wall_time_sec_mean,4.839267
confirm,fashion_wide,G8,val_acc_mean,41.666667
confirm,fashion_wide,G8,val_acc_std,2.269016
confirm,fashion_wide,G8,val_nll_mean,1.615937
confirm,fashion_wide,G8,val_nll_std,0.036692
confirm,fashion_wide,G8,wall_time_sec_mean,9.499333
confirm,fashion_wide,G5,val_acc_mean,46.31
confirm,fashion_wide,G5,val_acc_std,7.568494
confirm,fashion_wide,G5,val_nll_mean,1.525747
confirm,fashion_wide,G5,val_nll_std,0.149706
confirm,fashion_wide,G5,wall_time_sec_mean,10.038833
feasibility,mnist_compact,G8,execution_feasible,True
feasibility,mnist_compact,G8,optimization_feasible,True
feasibility,mnist_compact,G8,mean_val_nll,1.518089
feasibility,mnist_compact,G8,mean_val_acc,49.1533
feasibility,mnist_compact,G6,execution_feasible,True
feasibility,mnist_compact,G6,optimization_feasible,True
feasibility,mnist_compact,G6,mean_val_nll,1.642081
feasibility,mnist_compact,G6,mean_val_acc,45.9667
feasibility,mnist_wide,G8,execution_feasible,True
feasibility,mnist_wide,G8,optimization_feasible,True
feasibility,mnist_wide,G8,mean_val_nll,1.733351
feasibility,mnist_wide,G8,mean_val_acc,41.03
feasibility,mnist_wide,G5,execution_feasible,True
feasibility,mnist_wide,G5,optimization_feasible,True
feasibility,mnist_wide,G5,mean_val_nll,1.721259
feasibility,mnist_wide,G5,mean_val_acc,43.86
feasibility,fashion_compact,G8,execution_feasible,True
feasibility,fashion_compact,G8,optimization_feasible,True
feasibility,fashion_compact,G8,mean_val_nll,1.511217
feasibility,fashion_compact,G8,mean_val_acc,47.0033
feasibility,fashion_compact,G6,execution_feasible,True
feasibility,fashion_compact,G6,optimization_feasible,True
feasibility,fashion_compact,G6,mean_val_nll,1.584443
feasibility,fashion_compact,G6,mean_val_acc,45.6167
feasibility,fashion_wide,G8,execution_feasible,True
feasibility,fashion_wide,G8,optimization_feasible,True
feasibility,fashion_wide,G8,mean_val_nll,1.615937
feasibility,fashion_wide,G8,mean_val_acc,41.6667
feasibility,fashion_wide,G5,execution_feasible,True
feasibility,fashion_wide,G5,optimization_feasible,True
feasibility,fashion_wide,G5,mean_val_nll,1.525747
feasibility,fashion_wide,G5,mean_val_acc,46.31
1 section workload method metric value
2 protocol global all version HEAVY-TASK-PSO-V6 1.0.0
3 protocol global all official_test_data_loaded False
4 protocol global all official_test_evaluations 0
5 baseline mnist_compact untrained val_nll 2.325095
6 baseline mnist_compact untrained val_accuracy 6.65
7 baseline mnist_wide untrained val_nll 2.32458
8 baseline mnist_wide untrained val_accuracy 9.31
9 baseline fashion_compact untrained val_nll 2.320302
10 baseline fashion_compact untrained val_accuracy 8.08
11 baseline fashion_wide untrained val_nll 2.311222
12 baseline fashion_wide untrained val_accuracy 8.77
13 screen mnist_compact G0 val_nll 2.019988
14 screen mnist_compact G0 val_acc 33.32
15 screen mnist_compact G0 gbest_loss 2.014459
16 screen mnist_compact G0 gbest_acc 34.4
17 screen mnist_compact G0 throughput_sps 1106883.43
18 screen mnist_compact G5 val_nll 1.959743
19 screen mnist_compact G5 val_acc 33.95
20 screen mnist_compact G5 gbest_loss 1.971574
21 screen mnist_compact G5 gbest_acc 32.35
22 screen mnist_compact G5 throughput_sps 1139465.88
23 screen mnist_compact G6 val_nll 1.957307
24 screen mnist_compact G6 val_acc 35.11
25 screen mnist_compact G6 gbest_loss 1.941859
26 screen mnist_compact G6 gbest_acc 35.35
27 screen mnist_compact G6 throughput_sps 1340969.41
28 screen mnist_compact G8 val_nll 1.690442
29 screen mnist_compact G8 val_acc 43.88
30 screen mnist_compact G8 gbest_loss 1.675063
31 screen mnist_compact G8 gbest_acc 43.6
32 screen mnist_compact G8 throughput_sps 984615.38
33 screen mnist_wide G0 val_nll 2.088156
34 screen mnist_wide G0 val_acc 33.19
35 screen mnist_wide G0 gbest_loss 2.087951
36 screen mnist_wide G0 gbest_acc 32.45
37 screen mnist_wide G0 throughput_sps 595459.62
38 screen mnist_wide G5 val_nll 1.913566
39 screen mnist_wide G5 val_acc 39.87
40 screen mnist_wide G5 gbest_loss 1.924357
41 screen mnist_wide G5 gbest_acc 39.5
42 screen mnist_wide G5 throughput_sps 772511.47
43 screen mnist_wide G6 val_nll 2.017511
44 screen mnist_wide G6 val_acc 29.46
45 screen mnist_wide G6 gbest_loss 2.014518
46 screen mnist_wide G6 gbest_acc 27.8
47 screen mnist_wide G6 throughput_sps 777013.35
48 screen mnist_wide G8 val_nll 2.056341
49 screen mnist_wide G8 val_acc 20.43
50 screen mnist_wide G8 gbest_loss 2.062368
51 screen mnist_wide G8 gbest_acc 20.3
52 screen mnist_wide G8 throughput_sps 855005.34
53 screen fashion_compact G0 val_nll 1.969548
54 screen fashion_compact G0 val_acc 28.25
55 screen fashion_compact G0 gbest_loss 1.96974
56 screen fashion_compact G0 gbest_acc 28.15
57 screen fashion_compact G0 throughput_sps 1219977.13
58 screen fashion_compact G5 val_nll 1.892402
59 screen fashion_compact G5 val_acc 34.31
60 screen fashion_compact G5 gbest_loss 1.874629
61 screen fashion_compact G5 gbest_acc 36.6
62 screen fashion_compact G5 throughput_sps 1122281.97
63 screen fashion_compact G6 val_nll 1.824273
64 screen fashion_compact G6 val_acc 35.2
65 screen fashion_compact G6 gbest_loss 1.844914
66 screen fashion_compact G6 gbest_acc 33.5
67 screen fashion_compact G6 throughput_sps 1152322.65
68 screen fashion_compact G8 val_nll 1.88018
69 screen fashion_compact G8 val_acc 30.85
70 screen fashion_compact G8 gbest_loss 1.895269
71 screen fashion_compact G8 gbest_acc 28.85
72 screen fashion_compact G8 throughput_sps 1193139.45
73 screen fashion_wide G0 val_nll 1.981316
74 screen fashion_wide G0 val_acc 36.39
75 screen fashion_wide G0 gbest_loss 1.98436
76 screen fashion_wide G0 gbest_acc 36.05
77 screen fashion_wide G0 throughput_sps 540327.57
78 screen fashion_wide G5 val_nll 1.945862
79 screen fashion_wide G5 val_acc 31.97
80 screen fashion_wide G5 gbest_loss 1.938569
81 screen fashion_wide G5 gbest_acc 33.3
82 screen fashion_wide G5 throughput_sps 807401.18
83 screen fashion_wide G6 val_nll 1.975825
84 screen fashion_wide G6 val_acc 28.64
85 screen fashion_wide G6 gbest_loss 1.960139
86 screen fashion_wide G6 gbest_acc 28.1
87 screen fashion_wide G6 throughput_sps 735350.44
88 screen fashion_wide G8 val_nll 1.987923
89 screen fashion_wide G8 val_acc 28.69
90 screen fashion_wide G8 gbest_loss 1.976977
91 screen fashion_wide G8 gbest_acc 28.45
92 screen fashion_wide G8 throughput_sps 865800.87
93 confirm mnist_compact G8 val_acc_mean 49.153333
94 confirm mnist_compact G8 val_acc_std 1.320656
95 confirm mnist_compact G8 val_nll_mean 1.518089
96 confirm mnist_compact G8 val_nll_std 0.061523
97 confirm mnist_compact G8 wall_time_sec_mean 4.530167
98 confirm mnist_compact G6 val_acc_mean 45.966667
99 confirm mnist_compact G6 val_acc_std 3.545438
100 confirm mnist_compact G6 val_nll_mean 1.642081
101 confirm mnist_compact G6 val_nll_std 0.091398
102 confirm mnist_compact G6 wall_time_sec_mean 5.2706
103 confirm mnist_wide G8 val_acc_mean 41.03
104 confirm mnist_wide G8 val_acc_std 4.055083
105 confirm mnist_wide G8 val_nll_mean 1.733351
106 confirm mnist_wide G8 val_nll_std 0.084331
107 confirm mnist_wide G8 wall_time_sec_mean 9.4569
108 confirm mnist_wide G5 val_acc_mean 43.86
109 confirm mnist_wide G5 val_acc_std 3.512222
110 confirm mnist_wide G5 val_nll_mean 1.721259
111 confirm mnist_wide G5 val_nll_std 0.098656
112 confirm mnist_wide G5 wall_time_sec_mean 9.868467
113 confirm fashion_compact G8 val_acc_mean 47.003333
114 confirm fashion_compact G8 val_acc_std 5.2259
115 confirm fashion_compact G8 val_nll_mean 1.511217
116 confirm fashion_compact G8 val_nll_std 0.168628
117 confirm fashion_compact G8 wall_time_sec_mean 5.0146
118 confirm fashion_compact G6 val_acc_mean 45.616667
119 confirm fashion_compact G6 val_acc_std 3.585392
120 confirm fashion_compact G6 val_nll_mean 1.584443
121 confirm fashion_compact G6 val_nll_std 0.053077
122 confirm fashion_compact G6 wall_time_sec_mean 4.839267
123 confirm fashion_wide G8 val_acc_mean 41.666667
124 confirm fashion_wide G8 val_acc_std 2.269016
125 confirm fashion_wide G8 val_nll_mean 1.615937
126 confirm fashion_wide G8 val_nll_std 0.036692
127 confirm fashion_wide G8 wall_time_sec_mean 9.499333
128 confirm fashion_wide G5 val_acc_mean 46.31
129 confirm fashion_wide G5 val_acc_std 7.568494
130 confirm fashion_wide G5 val_nll_mean 1.525747
131 confirm fashion_wide G5 val_nll_std 0.149706
132 confirm fashion_wide G5 wall_time_sec_mean 10.038833
133 feasibility mnist_compact G8 execution_feasible True
134 feasibility mnist_compact G8 optimization_feasible True
135 feasibility mnist_compact G8 mean_val_nll 1.518089
136 feasibility mnist_compact G8 mean_val_acc 49.1533
137 feasibility mnist_compact G6 execution_feasible True
138 feasibility mnist_compact G6 optimization_feasible True
139 feasibility mnist_compact G6 mean_val_nll 1.642081
140 feasibility mnist_compact G6 mean_val_acc 45.9667
141 feasibility mnist_wide G8 execution_feasible True
142 feasibility mnist_wide G8 optimization_feasible True
143 feasibility mnist_wide G8 mean_val_nll 1.733351
144 feasibility mnist_wide G8 mean_val_acc 41.03
145 feasibility mnist_wide G5 execution_feasible True
146 feasibility mnist_wide G5 optimization_feasible True
147 feasibility mnist_wide G5 mean_val_nll 1.721259
148 feasibility mnist_wide G5 mean_val_acc 43.86
149 feasibility fashion_compact G8 execution_feasible True
150 feasibility fashion_compact G8 optimization_feasible True
151 feasibility fashion_compact G8 mean_val_nll 1.511217
152 feasibility fashion_compact G8 mean_val_acc 47.0033
153 feasibility fashion_compact G6 execution_feasible True
154 feasibility fashion_compact G6 optimization_feasible True
155 feasibility fashion_compact G6 mean_val_nll 1.584443
156 feasibility fashion_compact G6 mean_val_acc 45.6167
157 feasibility fashion_wide G8 execution_feasible True
158 feasibility fashion_wide G8 optimization_feasible True
159 feasibility fashion_wide G8 mean_val_nll 1.615937
160 feasibility fashion_wide G8 mean_val_acc 41.6667
161 feasibility fashion_wide G5 execution_feasible True
162 feasibility fashion_wide G5 optimization_feasible True
163 feasibility fashion_wide G5 mean_val_nll 1.525747
164 feasibility fashion_wide G5 mean_val_acc 46.31