Spatial transform

[skaae]

I created a spatial transformer layer example.

It uses some cluttered MNIST data created with deepminds generator, which i have converted to numpy format. I didn't relly get where we were supposed to put data :)

[skaae]

With Lasagne/Lasagne#361 merged I think this nearly ready.

@ebenolson: Can you help me figure out where to put the data?

[ebenolson]

@skaae I think I emailed you access directions for S3? If you have trouble with that let me know, or feel free to put them up on dropbox or something else like that.

Sorry that I haven't had time to test this out yet, but I'll try to in the next couple days, or just tell me when you think it's ready and I'll merge.

[skaae]

Thanks @ebenolson, I think it's in S3 now. Can one of you try if it works?

[f0k]

Oops, commented on the commit diff instead of the PR diff.

[skaae]

Thanks for the comments. The -N setting is usefull. The npz file seems to be compressed already. At least the filesize does not change if I use savez_compressed and the file is ~1.0gb if i use savez.

[skaae]

I removed the wrong wget output and added the -N setting.

[f0k]

The npz file seems to be compressed already. At least the filesize does not change if I use savez_compressed and the file is ~1.0gb if i use savez.

That's pretty good evidence then ;) I didn't download the file to check, to not stress Eben's S3 budget.

I removed the wrong wget output and added the -N setting.

Cool! Unfortunately, all outputs seem to be gone now, including the print statements and plots...

[skaae]

ups, that sucks. I'll have to run the model again then.

[f0k]

I'll have to run the model again then.

You can just copy all the outputs from your previous commit. git reflog should tell you the commit ID, and chances are that it's even still on github. (Or in the email notifications sent by github for my comments.)

Let me know if you need any help.

[f0k]

Actually, the easiest might be to git reset --hard HEAD@{1} to get back to your previous version, then remove the wget output by hand and add the -N by hand (i.e., by editing the notebook file in a text editor). Afterwards, you can amend and force-push again.

[skaae]

cool thanks. reflog is nice :)

The plots are back.

[f0k]

That's a myth, though. If you use the standard layers, the speed should be exactly the same, as Theano will use cuDNN both for convolution and pooling if it's available. It only makes a difference if you use strided pooling / convolution, because Theano may introduce a suboptimal gradient then (this will be fixed in future). Can you remove those checks to not spread the myth? (Again, it may be easiest to edit the file by hand. Note that you will also have to remove the Using DNN layers\n output above in this case.)

[skaae]

I removed the try/except

[f0k]

👍 Looks good to merge! (Didn't run it myself, though.)

[benanne]

This looks awesome :)

[ebenolson]

Very neat, thank you!