Sep-25-2019, 09:03 AM
Hi!
I am designing a Neural Network for a classification of time series. I have 4 classes of functions and around 50000 samples for each class. The functions have a length of about 500 points normalized so that max(abs(f(t)))=1. The 4 classes are roughly speaking 'continuous', 'discontinuous', 'in-between' and 'trash'.
What Neural Network design would you use for this particular problem?
The design I'm using currently is a CNN.
Thank you for your help!
I am designing a Neural Network for a classification of time series. I have 4 classes of functions and around 50000 samples for each class. The functions have a length of about 500 points normalized so that max(abs(f(t)))=1. The 4 classes are roughly speaking 'continuous', 'discontinuous', 'in-between' and 'trash'.
What Neural Network design would you use for this particular problem?
The design I'm using currently is a CNN.
model = Sequential() model.add(Conv1D(16, 8,input_shape=(500,1))) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(Conv1D(8, 5)) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(Conv1D(8, 3)) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(GlobalAveragePooling1D()) model.add(Dense(4, activation='softmax')) model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])I'm open to all suggestions. The Network is not performing very well (85% accuracy).
Thank you for your help!