Ability to change offline dataset sampling rate dynamically during training

We want to change sampling probability array(numpy) during training dynamically. Currently this array is fixed. In training we may want to use more data from offline dataset at first but later increase probability of online data that’s collected.