I have a dataset on EEG, with this shape: (11,1158, 200) Where 11 is the number of EEG channel 1158 is the number of each task 200 is the time interval of each task for example, if you plot a task, you'll get (Note that the data is normalized): This task represents a task with a class. (for example seeing a picture from class 2, and total number of classes in my dataset is 5). Now I converted my array to this shape: (1158, 200, 11) So that the model could differentiate each task. this is the model that I used: opt = keras.optimizers.Adam(learning_rate=1e-4) model = Sequential() model.add(Conv1D(filters=128, kernel_size=64, activation='relu', input_shape=(200, 11))) model.add(Conv1D(filters=64, kernel_size=8, activation='relu')) model.add(Dropout(0.5)) model.add(MaxPooling1D(pool_size=2)) model.add(Flatten()) model.add(Dense(100, activation='relu')) model.add(Dense(5, activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy']) model.fit(x_train, y_train, validation_data=(x_valid,y_valid), epochs=50, batch_size=16) I tried many different hyper-parameters, but all my results are somewhat like this: Epoch 50/50 58/58 [==============================] - 0s 5ms/step - loss: 0.1281 - accuracy: 0.9946 - val_loss: 2.7850 - val_accuracy: 0.1897 The training accuracy is high, but validation accuracy is between 20% to 25% (100/5 = 20 where 5 is the number of classes); Which basically means the model predicts something random. Is my approach wrong? If so, how should I solve this problem? Continue reading...
