SPSA in PyTorch

169 views
Skip to first unread message

Cassidy Laidlaw

unread,
Jul 11, 2019, 11:28:58 AM7/11/19
to cleverhans dev
I'm planning to work today on implementing SPSA in PyTorch under cleverhans.future.torch. I just wanted to put it here in case anybody else is/has been working on it

Nicolas Papernot

unread,
Jul 11, 2019, 11:34:42 AM7/11/19
to Cassidy Laidlaw, cleverhans dev
Thanks for your contribution, looking forward to the PR.

On Thu, Jul 11, 2019 at 8:28 AM Cassidy Laidlaw <cassi...@gmail.com> wrote:
I'm planning to work today on implementing SPSA in PyTorch under cleverhans.future.torch. I just wanted to put it here in case anybody else is/has been working on it

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-de...@googlegroups.com.
To post to this group, send email to cleverh...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/98df3518-3aff-4a8b-a55b-beee0a4656cd%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Cassidy Laidlaw

unread,
Jul 11, 2019, 3:05:04 PM7/11/19
to cleverhans dev
Just opened a PR!

Also, I think I discovered an issue with the existing (TensorFlow) implementations of SPSA in the library, which in normal usage cause the gradient to be underestimated by a factor of 100. Should I reach out to Jonathan Uesato, who authored the paper and implementation, or just fix the issue?

On Thursday, July 11, 2019 at 11:34:42 AM UTC-4, Nicolas Papernot wrote:
Thanks for your contribution, looking forward to the PR.

On Thu, Jul 11, 2019 at 8:28 AM Cassidy Laidlaw <cassi...@gmail.com> wrote:
I'm planning to work today on implementing SPSA in PyTorch under cleverhans.future.torch. I just wanted to put it here in case anybody else is/has been working on it

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-dev+unsubscribe@googlegroups.com.
To post to this group, send email to cleverhans-dev@googlegroups.com.

Nicolas Papernot

unread,
Jul 11, 2019, 3:15:22 PM7/11/19
to Cassidy Laidlaw, cleverhans dev
Can you raise an issue in the tracker and tag Jonathan?

On Thu, Jul 11, 2019 at 12:05 PM Cassidy Laidlaw <cassi...@gmail.com> wrote:
Just opened a PR!

Also, I think I discovered an issue with the existing (TensorFlow) implementations of SPSA in the library, which in normal usage cause the gradient to be underestimated by a factor of 100. Should I reach out to Jonathan Uesato, who authored the paper and implementation, or just fix the issue?

On Thursday, July 11, 2019 at 11:34:42 AM UTC-4, Nicolas Papernot wrote:
Thanks for your contribution, looking forward to the PR.

On Thu, Jul 11, 2019 at 8:28 AM Cassidy Laidlaw <cassi...@gmail.com> wrote:
I'm planning to work today on implementing SPSA in PyTorch under cleverhans.future.torch. I just wanted to put it here in case anybody else is/has been working on it

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-de...@googlegroups.com.
To post to this group, send email to cleverh...@googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-de...@googlegroups.com.
To post to this group, send email to cleverh...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/31d9e8eb-48a6-41c0-9d0d-d15b2e16fe71%40googlegroups.com.
Reply all
Reply to author
Forward
0 new messages