Brain-node-classification

Modified from https://github.com/Charliefff/Brain-node-classification/tree/main

Modified Parts

1. load_data.py → load_data_v2.py

• Use full 0–499 timesteps per trial (previously 0–374).
• Added wavelet transform for feature extraction.
• Changed data split from train / val / test to train / val only — planning to use n-fold cross-validation.
• Apply z-score normalization on validation data using training set mean and std for a more realistic evaluation setting.

---

2. model.py

• Added flatConv_waveletEEGNet architecture.
• Currently only tested EEGNet, waveletEEGNet, and flatConv_waveletEEGNet.
• flatConv_waveletEEGNet_v2 (which performs convolution over both band and time dimensions) is commented out for now —
  its current input shape (batch, 1, channel, 1) is incompatible with EEGNet’s temporal convolution structure.
  (Left for reference and possible future modification.)

---

3. main.py

• Replaced TensorBoard with standard logging.
• Introduced a workspace directory to store log files and model checkpoints.
• Removed test phase, since only train / val are used at this stage.

---

4. run_batch.py

• Added this script to automate running main.py across multiple parameter configurations.

Usage

1. Set Up the Environment

pip install -r requirements.txt

2. Modify Configuration

Before training, update config.yaml with appropriate hyperparameters.

3. Train the Model

(1) Train the CNN Model

cd conv_based
python main.py --config config.yaml

4. See Training Result

• Training results are saved under the directory:
  {log_dir}/{exp_name}_{time}
• Both log_dir and exp_name can be configured in config.yaml.
• Each run creates a unique timestamped subfolder to store:
  • Training logs
  • Model checkpoints
  • config file