search

millardjn / rusty_sr

Deep learning superresolution in pure rust
203 stars 21 forks source link

Could you share the parameters used for the README images? #2

Closed anowell closed 7 years ago

anowell commented 7 years ago

I haven't figured out how to get comparable results for the images on the README. If I run butterfly_nn.png with the imagenet parameters, I get a high-res (2304x1539) image that basically looks like a stretched image:

before-and-after

I've tried a few combinations of parameters with the different images from your README and seem to minor variations on the same results. I feel like I'm overlooking something. Any insights?

millardjn commented 7 years ago

Hi! The _nn files are nearest neighbour upscalings of the low resolution files which exist just for ease of formatting the readme. You can downscale those or the original set14 images by a factor of 3 to get the low resolution image back. Other than that, all images are default settings, except the cartoon on which the -p anime setting was used.

Downscaling for testing should be done with the -d flag or a program that supports linear RGB downscaling, for best results. Downscaling the _rs files isn't a good idea as the neural networks aren't perfectly idempotent.

millardjn commented 7 years ago

It seems reproducing the README examples is a common desire so I've committed the low resolution versions of the example images: https://github.com/millardjn/rusty_sr/issues/4.

millardjn commented 7 years ago

I hope the Algorithmia API goes well, feel free to contact me if you want any help with this, Alumina, or performance tweaking matrixmultiply_mt. Closing this for now.

anowell commented 7 years ago

The Algorithmia API implementation went well - it only took about an hour to port (you should recognize most of this code). Unfortunately, I uncovered a platform bug that is blocking me from publishing it as an API, but the fix is queued up for Monday. So stay tuned. :-)

And if you have any interest in owning/maintaining it on Algorithmia, just let me know. I'd be happy to hand it over. :-)

anowell commented 7 years ago

(oh. and yeah, your clarifications about the README images were exactly what I needed to know. 👍

millardjn commented 7 years ago

Thanks, that would be awesome! I've also got some blind deconvolution work which I think would be a good fit for the platform. I'll have a look around the dev api.