Back to projects

EmoTCN-Attn — Multi-Scale TCN + Attention for EEG Emotion Recognition

View resource →

  • PyTorch
  • EEG
  • emotion recognition
  • attention

What it is

A multi-scale Temporal Convolutional Network combined with multi-head self-attention for EEG-based emotion recognition, evaluated with Leave-One-Subject-Out cross-validation on SEED and SEED-V (94.3% / 91.2%).

The repo is notable for its ablation studies: it quantifies the contribution of attention (~38%), kernel-size diversity, TCN depth, dilation rates, and frequency-band selection — plus a full preprocessing pipeline (band-pass filtering, ICA artifact removal, normalization, augmentation). A ~2.3M-parameter model with 15 ms inference latency.

Stack

PyTorch · multi-scale TCNs · multi-head attention · ICA preprocessing · AdamW + OneCycleLR

Links