My suggestion:
 1. Shuffle the whole data set.
 2. Divide into training and test data.
 3. Apply SMOTE on training data.
 4. Follow other steps of Model training, without manipulating the data at this stage.
 5. Evaluate model on the actual test data.