In the demo code https://github.com/h2oai/sparkling-water/blob/master/py/examples/notebooks/TensorFlowDeepLearning.ipynb
I can more or less make out what the code is doing. My question is what is the advantage in creating the h2o model at the end for running it repeatedly. How is that better than just deploy the model in tensorflow?
You are basically training over 1 more pass of the data, which means the model has double the amount of steps to learn. Remember that any gradient descent sort of deep learning are progressing towards a local minimum in steps, so if you run 1 pass of the full data, it might not have arrive at the optimal point yet, so you might get better loss after more training.
