docs: update profiler tutorial

[anirudh161]

• Motivation for features / changes:

1. Saved the notebooks with outputs enabled (and private outputs disabled) as a workaround to prevent broken iframes rendering on tensorflow.org
2. Made some minor text changes and added a link to the Dev Summit 2020 talk

[review-notebook-app]

Check out this pull request on 

You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB.

[wchargin]

For context, this supersedes #3441.

[wchargin]

As listed in the check failure, please run this through nbfmt.py to
fix formatting errors and make the patch easier to review. The patch as
written is 240 lines, but is only 30 lines after formatting.

[wchargin]

Thanks! The prose changes look good to me. Was the broken-iframe issue
resolved separately? I don’t see any outputs added in this change, as
suggested by the PR description.

[anirudh161]

That's weird. The outputs were there before I ran nbfmt.py on the notebook. Let me look into it and submit another commit

[wchargin]

As I mentioned on #3441, make sure that private_outputs: false is
set in the notebook to preserve the contents. If you have a notebook
with output cells, you can set that easily by running nbfmt.py on the
notebook and passing --preserve_outputs. This will write the option
into the notebook so that future nbfmt.py runs know not to remove
them, even when the flag is not passed.

It doesn’t look like private_outputs: false was set at any point in
this change, but it also doesn’t look like the outputs were saved. The
“Formatted notebook with nbfmt” commit has a delta of 111 insertions and
110 deletions (i.e., just sorted keys and fixed trailing newline), and
you can see from the diff that it didn’t actually remove any outputs
(they were all outputs: [] already).

[anirudh161]

I was so sure I saved the outputs then! Guess I didn't after all :) I added another commit with outputs saved

[wchargin]

The TensorBoard output cells in this notebook don’t have cache entries
for the profile dashboard. When I open this notebook in Colab and
navigate to the TensorBoard cells without running them, I can use the
scalars dashboard, but the profile dashboard just shows “Not found”.

You can fix this by making sure to load the profile dashboard and select
the correct trace after running all cells in the notebook. The way that
this works is that any requests made during execution are cached when
you save the notebook, so you have to actually load the dashboard and
its data for the requests to persist.

This seems kind of important to me given that it takes multiple minutes
just to run the notebook to the point where it can be interactive, but
I’ll leave it up to you. Shall I merge this as is, or would you like to
fix the cache?

[anirudh161]

That's a good point. I fixed the cache entries and added a new commit

[wchargin]

Awesome; thanks! Tested; this works for me both before and after running
the notebook.

I’ll go ahead and merge this now.