WaveMind's EEG encoder training module for aligning EEG signals with CLIP semantic space through contrastive learning and supervised classification.

# Basic CLIP training (project root is auto-detected) python EEG_Encoder/run_CLIPtraining.py --config-name=train_atms

Pure CLIP Training (default):

python EEG_Encoder/run_CLIPtraining.py --config-name=train_atms

Joint CLIP + Classifier (lambda=0.5):

python EEG_Encoder/run_CLIPtraining.py --config-name=train_classifier

Pure Classifier Training (lambda=0.0):

python EEG_Encoder/run_CLIPtraining.py --config-name=train_classifier_only

All presets are in EEG_Encoder/examples/. See YAML files for detailed parameters:

• train_atms.yaml - CLIP training (default)
• train_classifier.yaml - Joint CLIP + Classifier (lambda=0.5)
• train_classifier_only.yaml - Pure supervised classifier (lambda=0.0)
• eval_sd.yaml - Subject-dependent evaluation
• eval_si.yaml - Subject-independent evaluation
• advanced_shm.yaml - High-performance with shared memory
• base.yaml - Full parameter reference

# Train on specific dataset python EEG_Encoder/run_CLIPtraining.py --config-name=train_atms experiment.datasets=[TUEV] # Use different model python EEG_Encoder/run_CLIPtraining.py --config-name=train_atms experiment.models=[EEGITNet] # Multi-GPU training python EEG_Encoder/run_CLIPtraining.py --config-name=train_atms experiment.gpu_number=['0','1','2'] # Adjust classifier lambda (80% CLIP, 20% classifier) python EEG_Encoder/run_CLIPtraining.py --config-name=train_classifier classifier.lambda_clip=0.8 # Evaluation with checkpoint python EEG_Encoder/run_CLIPtraining.py --config-name=eval_sd \ advanced.model_checkpoint_name=/path/to/model.pth # Advanced: shared memory + dynamic sampling python EEG_Encoder/run_CLIPtraining.py --config-name=advanced_shm \ training.DEFAULT_NUM_WORKERS=64

ATMSmodify (default), ATMS, NICE, EEGITNet, EEGConformer, ShallowFBCSPNet, CBraMod, NeuroLM-B/L, MLP

See examples/base.yaml for complete list and model-specific parameters.

Checkpoints saved to: EEG_Encoder/Resource/Checkpoint/ALL/

Naming convention: {Model}_{Dataset}_{Timestamp}.pth

• CLIP-only: No suffix
• Pure classifier: _CLS suffix
• Joint training: _JOINT{lambda} suffix (e.g., _JOINT50 for lambda=0.5)

• Configuration details: See YAML files in EEG_Encoder/examples/
• Project overview: CLAUDE.md in project root
• Full pipeline: WaveMind documentation