Reduce CUDA requirement

[eds89]

Hello,
Would there be any easy way to get the experiment running on CUDA 6.5? I stumbled onto the fact it requires CUDA 9.0. This is unfair as graphics cards which such support are the most expensive ones.

[felipecode]

Yes, I also think it is unfair !

I will eventually work on that. I think it shouldn't be hard to make it
work with tensorflow CPU. However downgrading to a tensorflow version
compatible with cuda version 6.5 might be tricky.

[eds89]

Maybe looking for tensorflow-gpu to cpu wrappers, if such a thing exists?

…

On 3 May 2018 at 16:04, Felipe Codevilla ***@***.***> wrote: Yes, I also think it is unfair ! I will eventually work on that. I think it shouldn't be hard to make it work with tensorflow CPU. However downgrading to a tensorflow version compatible with cuda version 6.5 might be tricky. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#29 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQUNuUcISpDOAc_-hghItFPNezVOhmcHks5tuxyEgaJpZM4Tu44G> .

[eds89]

Hi, Isn't it simply a matter of changing line 36 in '\imitation_learning.py'? On 3 May 2018 at 16:06, Eduardo dos Santos Work < [email protected]> wrote:

…

Maybe looking for tensorflow-gpu to cpu wrappers, if such a thing exists? On 3 May 2018 at 16:04, Felipe Codevilla ***@***.***> wrote: > Yes, I also think it is unfair ! > > I will eventually work on that. I think it shouldn't be hard to make it > work with tensorflow CPU. However downgrading to a tensorflow version > compatible with cuda version 6.5 might be tricky. > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#29 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AQUNuUcISpDOAc_-hghItFPNezVOhmcHks5tuxyEgaJpZM4Tu44G> > . >

[eds89]

Hi. Any updates on this feature request? I'm eager to repeat the imitation learning requirement. Unfortunately, can't do so at the moment due to hardware incompatibility.

[mvpcom]

Honestly, I believe if you reduce the computational power to CPU then it will not be possible to experience the same experiment. Using a powerful GPU like 1080Ti is one of the reasons for performing well at least with that FPS.

[eds89]

This is also a very expensive graphics card and compromises the reproducibility of the experiment.

…

On 7 June 2018 at 12:52, Mojtaba Valipour ***@***.***> wrote: Honestly, I believe if you reduce the computational power to CPU then it will not be possible to experience the same experiment. Using a powerful GPU like 1080Ti is one of the reasons for performing well at least with that FPS. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#29 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQUNuVVy5B3gsgdB819WTJ7BQzRzgOGgks5t6RQHgaJpZM4Tu44G> .

[eds89]

Hello, Would you have any idea how I may downgrade to TensorFlow-CPU or to a lower CUDA version? I'm hardly in need to try the conditional imitation learning experiments myself. Regards, On 7 June 2018 at 14:08, Eduardo dos Santos Work < [email protected]> wrote:

…

This is also a very expensive graphics card and compromises the reproducibility of the experiment. On 7 June 2018 at 12:52, Mojtaba Valipour ***@***.***> wrote: > Honestly, I believe if you reduce the computational power to CPU then it > will not be possible to experience the same experiment. Using a powerful > GPU like 1080Ti is one of the reasons for performing well at least with > that FPS. > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#29 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AQUNuVVy5B3gsgdB819WTJ7BQzRzgOGgks5t6RQHgaJpZM4Tu44G> > . >

[felipecode]

I am bit short on time to do that. But this should be quite straightforward.
Just install tensorflow-cpu or other versions. Then you should probably remove
the gpu parameters related things on the code.

[eds89]

Ok. Do you think the whole code should be restructured in order to migrate to tensorflow-cpu? Can you think of anything that would not be work by simply changing parameters?

…

On 11 June 2018 at 06:34, Felipe Codevilla ***@***.***> wrote: I am bit short on time to do that. But this should be quite straightforward. Just install tensorflow-cpu or other versions. Then you should probably remove the gpu parameters related things on the code. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#29 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQUNuQgo0ElRD-SR_KZuwJVTHdnzOrTuks5t7gGAgaJpZM4Tu44G> .