Spatial Transformer Network support in Torch with CuDNNv5?

[Abhishek Shivkumar]

Hi,

  I wanted to know if CuDNNv5 spatial transformer networks would be supported in any upcoming version of Torch?

In the current version of Torch, is there a way to make use of the spatial transformer network module present in CuDNN v5?

Thanks

Abhishek S

[soumith]

Yes, the bindings for the SpatialTransformer layer should be written shortly after May 20th. Until then, you can use https://github.com/qassemoquab/stnbhwd

--
You received this message because you are subscribed to the Google Groups "torch7" group.
To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at https://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[Abhishek Shivkumar]

Hi Soumith,

  

   Thanks for this reference. I was able to integrate it into my code but I am stuck with something. How can I fix the scale parameter in the Spatial Transformer Module to 50% (0.5) and just learn the two translation parameters?  I obviously cannot set the scale also as a learnable parameter because it would then change during training.

Can you please guide me to initialize the scale to 50% (0.5) and keep it constant during training while just the translation params change please?

Thanks
Abhishek S

On Tuesday, May 17, 2016 at 5:43:39 PM UTC+1, smth chntla wrote:

Yes, the bindings for the SpatialTransformer layer should be written shortly after May 20th. Until then, you can use https://github.com/qassemoquab/stnbhwd

[Abhishek Shivkumar]

Hi,

  I had to write my own Torch layer to fix the scale to 50%. Is there any update on the Spatial Transformer Network wrapper for cudnn v5?

[Daniel Galvez]

As someone behind on wrapping CUDNN inside another toolkit, don't be afraid to do this yourself (though you should see if there's an issue for this in torch.cudnn and ask about its closeness to completion). I would be happy if someone took over the task for me :P

Wrapping CUDNN is not hard, per-se, if you know C. There is just a lot of code you have to write to set tensor descriptors and delete them after you've used them, since C is so low-level.

[tanzila rahman]

Hi @Abhishek how do you fix the scale parameter in STN? I am alos stuck in this point. Can you please help me regarding this?

[Anurag Ranjan]

Have the bindings been written yet?

On Tuesday, May 17, 2016 at 9:43:39 AM UTC-7, smth chntla wrote:

Yes, the bindings for the SpatialTransformer layer should be written shortly after May 20th. Until then, you can use https://github.com/qassemoquab/stnbhwd

[Arjun Jain]

Hi Daniel!

I too would like to use the torch spatial transformer network bindings. Do you have any ETA? I am also happy to take the work over from you, I have contributed torch modules in the past. Do let me know if you wanna go this route.

Regards,

Arjun

[Daniel Galvez]

Hi Arjun,

Sorry if there is a misunderstanding, but I never was working on wrapping CUDNN's spatial transformer module in torch7.

Sincerely,

Daniel

--
You received this message because you are subscribed to a topic in the Google Groups "torch7" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/torch7/jPf9GTFMUeo/unsubscribe.
To unsubscribe from this group and all its topics, send an email to torch7+unsubscribe@googlegroups.com.

To post to this group, send email to tor...@googlegroups.com.
Visit this group at https://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

--

Daniel Galvez

[Arjun Jain]

Ah, sorry for the misunderstanding, my bad. 

On Wed, May 31, 2017 at 11:55 AM, torch7 on behalf of Daniel Galvez <tor...@googlegroups.com> wrote:

Hi Arjun,

Sorry if there is a misunderstanding, but I never was working on wrapping CUDNN's spatial transformer module in torch7.

Sincerely,

Daniel

To unsubscribe from this group and stop receiving emails from it, send an email to torch7+unsubscribe@googlegroups.com.