Implementation of Convolutional Spatial Transformer and Siamese network

[Sentient07]

We can have an example of a Siamese network (similar to Sander Dieleman's suggestion ) and convolutional spatial transformer as discussed here. From my understanding, the convolutional spatial transformer applies the ST on the feature, patchwise with distinct affine transformation parameters. I have implemented this very recently and will open a PR if you are fine with having these examples in the Recipes @f0k .

PS : I have a few PRs and issues open in Lasagne which I'm yet to resolve. The past two weeks were a bit hectic for me and I will try to resolve them by this weekend. Apologies

[f0k]

We can have an example of a Siamese network

Yes, that would be useful. I'm going to expand the documentation sometime and plan to include a hint at how to do it, but a Recipe would be good as well. Maybe some silly toy example such as embedding MNIST digits into a 2D space such that items of the same class are closer together. Would give a nice notebook.

and convolutional spatial transformer

That will be more work, but feel free to have a go.

I have a few PRs and issues open in Lasagne which I'm yet to resolve. The past two weeks were a bit hectic for me and I will try to resolve them by this weekend.

No worries, I'm quite busy myself currently!

[Sentient07]

I'm going to expand the documentation sometime

Expand the documentation as in?

Would give a nice notebook.

I think we can try and have as many notebooks as possible of some of the breakthrough papers at a top-level conferences. But again, judging a paper as breakthrough or not is subjective and i'm not sure how it'd work

That will be more work, but feel free to have a go.

I have done an implementation. I will share here soon. But do you think it would be better to have it Lasagne itself?

[f0k]

Expand the documentation as in?

The general Lasagne documentation at lasagne.readthedocs.io, not the module/API documentation. The MNIST example leaves a lot of open questions.

But again, judging a paper as breakthrough or not is subjective and i'm not sure how it'd work

We only have limited resources, we cannot implement everything -- people will implement the papers they are interested in or would like to expand on. So it'll always be subjective, but that's okay.

But do you think it would be better to have it Lasagne itself?

Starting with a Recipe is always a good idea, that's a lower hurdle and we can migrate parts to Lasagne at any time.