Improve an existing Keras Code

I need help in adding more functionality to my existing Keras code: 1) I need the output of Keras predictions to be indexed and stored on to a new column in the data frame from where the test sample was taken. The row numbers of the prediction stored in the dataframe should be similar to the test set row numbers. 2) Improvise the Keras code by adding multi-gpu and parallel processing to the training and prediction part of the code 3) Create a functionality in the existing model using keras/tensorflow to train the model batch-wise ( As my entire data is not available at the time of initially training the model, I will need to train the model as and when new data comes).