# 8-channel Shallow Transformation Recognition SSL

This experiment removes masked reconstruction entirely. A ShallowConvNet
encoder predicts five synthetic transformation classes from continuous EEG,
then transfers to subject-specific left/right motor-imagery classification.

## Commands

```bash
python run_transform_ssl.py smoke-test --device cuda
python run_transform_ssl.py pretrain --device cuda
python run_transform_ssl.py finetune --device cuda
python run_transform_ssl.py report --device cpu
python run_transform_ssl.py all --device cuda
```

The validated `data_cache_1_40` from `SSL_8ch_ConvMAE_MI_CUDA_package` is
reused. The ten final subjects never participate in pretraining or checkpoint
selection. Final evaluation includes both all-trial and exactly 20+20
calibration settings with three seeds and five outer folds.