Training sgcls with ResNet ?

[rowanz]

Training sgcls with ResNet ?

[FrancoisPl]

Hi Rowan!

Thank you for your answer!
I've made a few changes to the code to use resnet 34 (I had memory issues with resnet 101) and made a Pull request with the changed code, if you're interested or want to give me some feedback?

Furthermore, on my end, the training takes a long time before it starts. Do you know why and did you simply use a smaller dataset when iterating on the code?

Thanks again!!

[rowanz]

hi Francois, thanks for the interest :) Unfortunately, I don't support ResNet anymore - for an apples to apples comparison with prior work, I redid everything using VGG. Most of the things should be the same, however, you may need to tweak the layer sizes, as well as how you incorporate the edge features (which use some additional convolutions). It's probably not too bad in theory, but for simplicity's sake the easiest thing to do would be to just use my VGG detector :)

[rowanz]

thanks again for the PR! My guess as for why it takes a while before training begins is that you have to initialize Resnet (or whatever CNN you're using), but I'm not sure.

[galsk87]

@FrancoisPl can you share your code?