Add 2D bilinear upsampling layer

[christopher-beckham]

Hi,

After being frustrated so many times with deconvolution artifacts, I decided to try out bilinear upsampling and it turns out this already existed in Theano, so I thought it would be reasonable to contribute this as a layer to Lasagne. I have tried this so far with some of my research involving pix2pix (conditional image-to-image translation with GANs), and the use of this + followed by a shape-preserving convolution (i.e., a conv with padding=same) gave me much better outputs than simply using deconv!

You can read more about it here:

http://distill.pub/2016/deconv-checkerboard/

Admittedly, my unit test test_get_output_for is a bit crude -- it simply checks the output shape, and not against a Numpy implementation of bilinear upsampling. Apart from that, everything seems to look ok. Feel free to offer suggestions, etc. I'd love to see this kind of layer upstream.

[f0k]

Thank you for the contribution and sorry for the late reply, I've been busy with my thesis.

gave me much better outputs than simply using deconv!

The easier / more efficient solution that popped up in between may be #862.

Still, bilinear upsampling would be nice. What about adding this as a different mode to Upscale2DLayer instead? We could have "bilinear" for all-in-one and "bilinear1d" for separate passes per dimension. It's nice that the complexity of creating the kernels and convolution expressions is hidden away in Theano.

[christopher-beckham]

Interesting, yes, I think it would be cleaner to have this as an option for Upscale2D. I'll have a look at that paper too.

[f0k]

Interesting, yes, I think it would be cleaner to have this as an option for Upscale2D.

Would you like to work on this?

[christopher-beckham]

Hi f0k, gahhh sorry for that one month delay... looks like we've both been swamped. Yes, absolutely, I'd be happy to. Let me try get back to you next week.