Thomas Verelst

@zzx0836 Hi! I changed some things in helper.py, but now it doesnt seem to run anymore (hangs on the first sample). If you want to take a look what I...

Hi, for me annotations are persistent when setting the "Attachments Location" to a Custom Path in the settings.

Hi, I did not find a fast and straightforward CUDA implementation of standard convolutions so I was not able to integrate the method. Standard convolutions are hard to implement efficiently...

Hi, thanks for your interest in our work! I suppose you are referring to the classification code. In classification, we use the same mask unit as the work we compare...

Hi, thanks for having a look at the code. I did not test dual-gpu training, and RN101 indeed takes quite some time on single GPU (~2 weeks). I did not...

I've pushed a new branch `multigpu`. I didn't test training accuracy yet, but it runs. I only had problems with gathering the output dict `meta`. I considered subclassing `DataParallel` to...

Thanks a lot, that fixed it.

Hi, I don't have a Colab version - I generally don't use Colab - but will take it into account for future versions.

Hi! I think you mean that it uses the straight-through version of the Gumbel-Softmax trick (hard version). I did not thoroughly ablate this, but my initial results indicated slightly better...

Some things that could help for convergence; - one of the gumbel-softmax papers shows that for a binary case, the temperature should be smaller or equal to 1 for convergence....