Leave-One-Subject-Out Cross-Validation (LOSV) for DGCNN Model in SEED Dataset

Pseudo code for the given code in English:

Define a function 'main_LOCV()':

• Set the directory path as 'D:/SEED/SEED/SEED_code/DE/session1/'.

• Get the list of files in the directory and store the number of files in 'sub_num'.

• Define the dimensions of the input tensor as 'xdim = [128, 62, 5]'.

• Set the value of 'k_adj' as 40, 'num_out' as 64, 'num_epochs' as 5, 'acc_mean' as 0 and 'acc_all' as an empty list.

• For each 'sub_i' in the range of 'sub_num':

  • Set the 'load_path' as the directory path concatenated with the file name.

  • Load the data and labels from the file using the 'load_DE_SEED' function.

  • Normalize the data using the 'zscore' function.

  • Initialize a 'DGCNN' model with 'xdim', 'k_adj' and 'num_out' and move it to the device.

  • Define the loss function as 'CrossEntropyLoss'.

  • Define the optimizer as 'Adam' with learning rate 0.001 and weight decay 0.0001.

  • Define the 'train_list' as a deep copy of 'file_list' and remove the current file from it.

  • Set 'train_num' as the length of 'train_list'.

  • For each 'train_i' in the range of 'train_num':

    • Set 'train_path' as the directory path concatenated with the file name.
    • Load the data and labels from the file using the 'load_DE_SEED' function.
    • Normalize the data using the 'zscore' function.
    • Concatenate the data and labels to 'data_train' and 'label_train' respectively.
    • Convert 'data_train' and 'label_train' to tensors.

  • Load the training and testing data using the 'load_dataloader' function.

  • Train the model using the 'train' function and get the best testing accuracy.

  • Update the value of 'acc_mean' as 'acc_mean' plus 'acc_test_best/sub_num'.

  • Append the value of 'acc_test_best' to 'acc_all'.

• Save the values of 'acc_all' and 'file_list' as a MAT file.

• Print the LOSV test accuracy, mean accuracy and standard deviation of accuracy from 'acc_all'.