Tensorboard: Work around `fromtimestamp(0)` bug in Windows Python 3.6.0…3.6.6

[wchargin]

See #1976.

[wchargin]

For the record, I’ve attempted to do this a few times and given up each
time, because I can’t find a good way to implement strftime("%s")
portably. Using (dt - EPOCH).total_seconds() (the current code) is not
correct because it incorporates daylight savings time…

>>> fts = datetime.datetime.fromtimestamp
>>> (fts(9971999) - fts(0)).total_seconds()
9971999.0
>>> (fts(9971999 + 1) - fts(0)).total_seconds()
9975600.0

In Python 3, datetime.timezone.utc looks like it should at least
provide the foundations to solve these issues, but without adding a
pytz dependency we don’t have a Python 2 equivalent, and the code to
backport is not trivial (I checked).

I might just throw up my hands and replace everything in Python memory
with bare ints-since-epoch. At least then we just have time.time() to
get the real value and str/int for serialization—no syscalls, no OS
errors. I guess I’m just not Dutch enough for this one. *sigh*

[stefanbschneider]

Actually, this was a quick and easy solution for me: https://github.com/tensorflow/tensorboard/issues/1976#issuecomment-471834147

pip install tensorboard==1.12.2

[wchargin]

@stefanbschneider: That downgrades you to a version of TensorBoard prior
to the notebook integration features. You should be able to safely
upgrade to TensorBoard 1.14, which has both the features and the fix.

[stefanbschneider]

Perfect, thanks!