Update settings.py

x 20 to training steps. 500K is too low for a large training set IMO. Not sure why train steps is an hparam at all, either should be endless or maybe dynamic based on training data size to match like... 5-10 epochs.

Update README.md

full tf-gpu instructions