Keras: theano is going to switch from cuda to gpuarray, are we going to support api changes?

[fayeshine]

WARNING (theano.sandbox.cuda): The cuda backend is deprecated and will be removed in the next release. Please switch to the gpuarray backend. You can get more information about how to switch at this URL:
https://github.com/Theano/Theano/wiki/Converting-to-the-new-gpu-back-end%28gpuarray%29

theano is going to switch from cuda to gpuarray, the api may have some changes, will we support this change?

[Dapid]

I believe it is already compatible. For example, Keras is using theano.tensor.nnet.conv2d and the only occurences of theano.sandbox.cuda.dnn are in instances of _old_batch_normalization and the like.

I am installing libgpuarray and will do some testing.

[Dapid]

I have a simple 3D CNN that works with the new backend. The runtime seems to be slightly longer ( < 10%), but I haven't activated any of the new optimisations.

[fayeshine]

@Dapid , How to activate new optimizations in keras?

[Dapid]

Through THEANO_FLAGS. The ones I looked at aren't activated by default because they take more memory. http://deeplearning.net/software/theano/library/tensor/nnet/conv.html

For example, THEANO_FLAGS=optimizer_including=conv3d_fft:convgrad3d_fft:convtransp3d_fft

You can also activate it per function. In Keras you can pass them as kwargs in model.compile.

[nouiz]

Those are old flags for the old back-end. They won't affect the speed on
the new back-end.

To compare the speed of both back-end, you should ignore the time in the
first call. Some compilation happen at that moment when we know the strides.

On Tue, Feb 7, 2017 at 10:56 AM Dapid [e-mail] wrote:

Through THEANO_FLAGS. The ones I looked at aren't activated by default
because they take more memory.
http://deeplearning.net/software/theano/library/tensor/nnet/conv.html

For example,
THEANO_FLAGS=optimizer_including=conv3d_fft:convgrad3d_fft:convtransp3d_fft

You can also activate it per function. In Keras you can pass them as
kwargs in model.compile.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/fchollet/keras/issues/5303#issuecomment-278042882,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AALC-7JvgeChISkVXQsplfgeRaOO04qPks5raJQPgaJpZM4L5HPY
.

[nouiz]

I made an issue to update this doc:
http://deeplearning.net/software/theano_versions/dev/library/tensor/nnet/conv.html

On Tue, Feb 7, 2017 at 11:36 AM Frédéric Bastien [e-mail]
wrote:

Those are old flags for the old back-end. They won't affect the speed on
the new back-end.

To compare the speed of both back-end, you should ignore the time in the
first call. Some compilation happen at that moment when we know the strides.

On Tue, Feb 7, 2017 at 10:56 AM Dapid [e-mail] wrote:

Through THEANO_FLAGS. The ones I looked at aren't activated by default
because they take more memory.
http://deeplearning.net/software/theano/library/tensor/nnet/conv.html

For example,
THEANO_FLAGS=optimizer_including=conv3d_fft:convgrad3d_fft:convtransp3d_fft

You can also activate it per function. In Keras you can pass them as
kwargs in model.compile.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/fchollet/keras/issues/5303#issuecomment-278042882,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AALC-7JvgeChISkVXQsplfgeRaOO04qPks5raJQPgaJpZM4L5HPY
.

[HenryJia]

Hmm, although do you find that you get a discrepancy between the old and the new backends?
I can't seem to get the new backend to work at all
https://github.com/Theano/Theano/issues/5475

[nouiz]

Just a follow up, the problem was a bad NVIDIA driver. Now inside
libgpuarray we black list it.

On Tue, Feb 7, 2017 at 6:51 PM Hengjian Jia [e-mail]
wrote:

Hmm, although do you find that you get a discrepancy between the old and
the new backends?
I can't seem to get the new backend to work at all
Theano/Theano#5475 https://github.com/Theano/Theano/issues/5475

—
You are receiving this because you commented.

Reply to this email directly, view it on GitHub
https://github.com/fchollet/keras/issues/5303#issuecomment-278183375,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AALC-2-kMzLICy8zeI-1ZJ8d_ulpyCE_ks5raQN0gaJpZM4L5HPY
.

[alyato]

hi @fayeshine @Dapid @nouiz I install the latest keras which is 2.0.4, it show me an error. Then i upgrade the theano to 0.9 dev , the new theano show me switch to new gpu backend.

ValueError: You are tring to use the old GPU back-end. It was removed from Theano

Does i switch the new gup backend ? or is there good idea? Thanks.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs, but feel free to re-open a closed issue if needed.