Scalability

@archydeberker Thank you very much for sharing this repository! I have looked into your code. As far as I understood each thread will have its session because of estimator is encapsulating the session. Consequently the graph will be copied. If the model is big like BERT, it will hurt cache performance of CPU/GPU. Pease let me know if I'm thinking correctly or not.