Switch to tensorflow, pluses and minuses:
- more difficult than keras
- batch normalization harder to implement correctly
- more flexible
This is a basic example, just takes training steps, but we don't know how well it is doing.
For this exercise, still need train/predict arg, but only train implmented
code: ex04_tf_train.py
- Computation Graph - note with Graph at bottom
- Explicitly creating variables
- No Batch norm yet, using bias adds
- Idea of session
- Initializing variables