輸入固定為 32 通道 × 400 samples；比較 SSL full fine-tune、linear probe 與 scratch。
資料快取
| dataset | subject_id | split | kept_windows | excluded_windows | left_trials | right_trials |
|---|---|---|---|---|---|---|
| pretrain | sub16 | train | 1164 | 0 | - | - |
| pretrain | sub16 | val | 125 | 0 | - | - |
| pretrain | sub17 | train | 1146 | 0 | - | - |
| pretrain | sub17 | val | 124 | 0 | - | - |
| pretrain | sub18 | train | 1147 | 0 | - | - |
| pretrain | sub18 | val | 123 | 0 | - | - |
| pretrain | sub19 | train | 1139 | 0 | - | - |
| pretrain | sub19 | val | 123 | 0 | - | - |
| pretrain | sub20 | train | 1170 | 0 | - | - |
| pretrain | sub20 | val | 126 | 0 | - | - |
| pretrain | sub21 | train | 1158 | 0 | - | - |
| pretrain | sub21 | val | 125 | 0 | - | - |
| pretrain | sub22 | train | 1136 | 0 | - | - |
| pretrain | sub22 | val | 123 | 0 | - | - |
| pretrain | sub23 | train | 1154 | 0 | - | - |
| pretrain | sub23 | val | 125 | 0 | - | - |
| pretrain | sub47 | train | 1116 | 0 | - | - |
| pretrain | sub47 | val | 120 | 0 | - | - |
| pretrain | sub49 | train | 1263 | 0 | - | - |
| pretrain | sub49 | val | 136 | 0 | - | - |
| labeled | sub1 | all | 203 | 0 | 100.0 | 103.0 |
| labeled | sub2 | all | 203 | 0 | 102.0 | 101.0 |
| labeled | sub3 | all | 200 | 0 | 100.0 | 100.0 |
| labeled | sub9 | all | 201 | 0 | 100.0 | 101.0 |
| labeled | sub10 | all | 200 | 0 | 100.0 | 100.0 |
| labeled | sub11 | all | 200 | 0 | 100.0 | 100.0 |
| labeled | sub12 | all | 201 | 0 | 100.0 | 101.0 |
| labeled | sub13 | all | 200 | 0 | 100.0 | 100.0 |
| labeled | sub14 | all | 200 | 0 | 100.0 | 100.0 |
| labeled | sub15 | all | 201 | 0 | 100.0 | 101.0 |
CPU smoke test
狀態：passed；遮罩數 [96, 96]；模型參數 3,126,032。
CUDA 正式訓練
尚未在 CUDA 設備完成 200-epoch 預訓練與 150 個微調 folds，因此本頁不呈現推測性的 accuracy。執行套件內的 python run_ssl.py all --device cuda 後，重新產生報告即可補齊全部圖表與 CSV。