TF2 was released and currently a new GPflow is being developed as well. TF2 is a major change and it seems to make a lot of things easier.

Looking at the development of GPflow, I believe we will be able to make a compatible version again. Unfortunately, we'll need to redesign and rethink the structure of GPflowOpt:

• We can delete a lot code, include the awful modelwrapper
• GPflow 2.0 seems to further decrease the uniformity in the way models deal with data and optimization of hyperparameters. This may still be improved but I don't think we should rely on this.

My idea now is to create an abstract BO class which implements the main flow and calls a bunch of abstract methods along the way. We provide an implementation class for vanilla BO which uses a GPR, more complex strategies are not supported out of the box. Some things to consider:

• using tf bijectors for scaling
• Rewrite domain as a tf.Module and use tf optimizers (need a transform for the bounds)?
• Include tf.summary for TensorBoard
• Use checkpointing to safeguard long running BO runs against crashes.