EEGSym
Open implementation and code from the publication "EEGSym: Overcoming Intersubject Variability in Motor Imagery Based BCIs With Deep Learning". [1]
[1] S. Pérez-Velasco, E. Santamaría-Vázquez, V. Martínez-Cagigal, D. Marcos-Martínez and R. Hornero, "EEGSym: Overcoming Inter-subject Variability in Motor Imagery Based BCIs with Deep Learning," in IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2022, doi: https://doi.org/10.1109/tnsre.2022.3186442
Architecture details
Datasets avaivability
Since the datasets used are publicly available you should download them from the original sources. Here you can find direct links to the articles and datasets:
Stieger2021 | Article | Dataset
Example of use:
from EEGSym_architecture import EEGSym
from EEGSym_DataAugmentation import trial_iterator
from tensorflow.keras.callbacks import EarlyStopping as kerasEarlyStopping
from tensorflow.keras.utils import to_categorical
import numpy as np
#%% Load signal and model
# Batch size used for fine-tuning
bs_EEGSym = 32 # bs_EEGSym = 256 for pre-training (To reduce compute time)
# Select if the pre-trained weight values on 5 datasets that include 280
# users are loaded from these electrode configurations:
# 8 Electrode configuration: ['F3', 'C3', 'P3', 'Cz', 'Pz', 'F4', 'C4', 'P4']
# 16 Electrode configuration: ['F7', 'F3', 'T7', 'C3', 'P7', 'P3', 'O1', 'Cz',
# 'Pz', 'F8', 'F4', 'T8', 'C4', 'P8', 'P4', 'O2']
pretrained = True # Parameter to load pre-trained weight values
# Select the number of channels of your application (8/16 for pretrained=True)
ncha = 8
# Load the X EEG features and Y labels of your choice in the following format:
# X = Features in the form of [examples, samples, ncha, 1]
examples = 1000
samples = 128 * 3 # frequency = 128 Hz * time= 3 seconds
# Random signal for demonstrative purposes
X = np.random.normal(loc=0, scale=1, size=[examples, samples, ncha, 1])
# Y = one_hot encoded labels, i.e., [examples, 1, 1]
Y = np.random.randint(low=0,high=2,size=examples)
Y = to_categorical(Y)
# Divide the features into training, validation and test to obtain:
# X_train, Y_train, X_validate, Y_validate, X_test, Y_test
X_train = X[:int(0.6*examples)]
Y_train = Y[:int(0.6*examples)]
X_validate = X[int(0.6*examples):int(0.8*examples)]
Y_validate = Y[int(0.6*examples):int(0.8*examples)]
X_test = X[int(0.8*examples):]
Y_test = Y[int(0.8*examples):]
#%% Initialize the model
hyperparameters = dict()
hyperparameters["ncha"] = ncha
hyperparameters["dropout_rate"] = 0.4
hyperparameters["activation"] = 'elu'
hyperparameters["n_classes"] = 2
hyperparameters["learning_rate"] = 0.0001 # 1e-3 for pretraining and 1e-4
# for fine-tuning
hyperparameters["fs"] = 128
hyperparameters["input_time"] = 3*1000
hyperparameters["scales_time"] = np.tile([125, 250, 500], 1)
hyperparameters['filters_per_branch'] = 24
hyperparameters['ch_lateral'] = int((ncha / 2) - 1) # 3/7 For 8/16 electrodes
# respectively. For the configurations present in the publication.
hyperparameters['residual'] = True
hyperparameters['symmetric'] = True
model = EEGSym(**hyperparameters)
model.summary()
# Select if the data augmentation is performed
augmentation = True # Parameter to activate or deactivate DA
# Load pre-trained weight values
if pretrained:
model.load_weights('EEGSym_pretrained_weights_{}_electrode.h5'.format(ncha))
# Early stopping
early_stopping = [(kerasEarlyStopping(mode='auto', monitor='val_loss',
min_delta=0.001, patience=25, verbose=1,
restore_best_weights=True))]
#%% OPTIONAL: Train the model
if pretrained:
for layer in model.layers[:-1]:
layer.trainable = False
fittedModel = model.fit(trial_iterator(X_train, Y_train,
batch_size=bs_EEGSym, shuffle=True,
augmentation=augmentation),
steps_per_epoch=X_train.shape[0] / bs_EEGSym,
epochs=500, validation_data=(X_validate, Y_validate),
callbacks=[early_stopping])
#%% Obtain the accuracies of the trained model
probs_test = model.predict(X_test)
pred_test = probs_test.argmax(axis=-1)
accuracy = (pred_test == Y_test.argmax(axis=-1))
#%% Optional: Store the weights
model.save_weights("\custom_path")