Open source code for paper Convolutional Dynamically Convergent Differential Neural Network for Brain Signal Classification
We use the datasets from the BCI Competition.
These datasets contain EEG data recorded from 64 electrodes placed on the scalp of subjects, sampled at a rate of 240 Hz. The original datasets are in MATLAB format. To simplify the data processing, we pre-converted the datasets into HDF5 format and segmented the EEG signals into samples of length 7794 (i.e. 32.475 seconds) per trial. The HDF5 files have been split into 40MB chunks and uploaded to the data directory.
Upon the first run, these split files will be automatically merged. Alternatively, you can manually merge them using the following command.
cat data/dataset.hdf5.part_* > data/dataset.hdf5Our preprocessing is fairly straightforward, consisting only of band-pass filtering and z-score normalization. For details, please refer to the preprocess.py file.
Please refer to the model.py.
We use the popular PyTorch framework for training. To utilize PyTorch's standard training workflow, we extend the torch.utils.data.dataset.Dataset class and use torch.utils.data.DataLoader to load the datasets. For details, please refer to the dataset.py file.
During training, we used the proposed ADCL optimizer. ADCL is not much different from other optimizers in Torch, except that it requires passing the loss through a closure during training. Additionally, we used tensorboard to track the changes in metrics during the training process. You can view this data using the following command.
tensorboard --logdir=runsThe training results will be saved in the result folder, and you can also manually specify the folder using the --result-path parameter.
For details on the training process, please refer to train.py.
python train.py --helpFor details, please refer to the grad_cam.py file.