mnist_dpsgd_tutorial is very slow

[gagandeep987123]

I am following the blog regarding the TF-Privacy which basically follows code from two files mnist_scratch.py and mnist_dpsgd_tutorial.py to give the overview.

While mnist_scratch.py is running fine, the other file with the given default setting for flag 'dpsgd' as True is running very slow in training. It is taking around 990 secs for 1 epoch when DP is enabled and 12 secs when it is disabled. Could you please help in checking the issue?

I have seen a similar kind of issues but could not find the resolution.

Thanks

[FredericW]

mnist_scratch supposes to be fast as it didn't use any dp optimizer. Unfortunately I have the exact issue like you but even worse! With standard noise_multiplier=1.1 and l2_nor=1 which means the stddev=1.1 it tooks almost half hour for one epoch on my side! So I can hardly test my modification... Wish someone else can answer our questions.