Fix #7

[rainwoodman]

Fix #7.

The SAM training accuracy is near 82.5%, increased from the quoted numbers on the home page.

The non-SAM training appears to terminated early at 5m30, ~ 70 epochs, and at a lower accuracy than the home page number.

[rainwoodman]

I am not sure if this ipynb file is clean -- I see some references to my own github repo.

[sayakpaul]

@rainwoodman the notebook is great. Just a few minor suggestions:

• In the cell that implements SAM class I guess we can skip the TensorFlow import. Also, I don't think we need to disable eager execution for the current setup. My understanding is that when train_step() is overridden and the underlying model is compiled it is going to run in graph mode already. However, please correct me if I am missing out on something here.
• You are welcome to edit the README numbers with the latest ones. This will give the users more confidence in using SAM I hope.

[rainwoodman]

You are right I forgot to remove the lines. Those were for debugging, when I checked that the before this change indeed we used to have gradients == sam_gradients.

[rainwoodman]

I removed the lines, and updated README. Please take a look, thanks!

[sayakpaul]

@rainwoodman thank you for the changes. Just one more minor comment.

[sayakpaul]

Did you also want to update this table with the final scores?

[rainwoodman]

Yes It was updated in the previous commit. ;)

[sayakpaul]

Oh I see. Anyway the current README should reflect the changes you made. Thank you so much.