Paul Balanca

@rlouf As we discussed on Slack, there is quite a bit of additional complexity to add to this PR to handle properly the support `select` condition appearing in lot of...

I have experimented on my side on Inception v2 model. The current regularisation was clearly hurting performance, as removing it and adding a few extras (label smoothing + aux logits),...

True, the aux logits is not part of Inception v2. I guess it only makes a minor difference at the end (0.2 if i remember well the paper). I also...

Hi, thanks :) Took me a bit of time to make it, but I guess it is worth it! The training script is not entirely stable. You may want to...

Do you mean rename it? You just need to download the checkpoint of the VGG-16 model, and use in the training command. Hopefully it should work!

Hello, Yes, the checkpoints are directly converted from the Caffe implementation. The training script is not yet as advanced as the latter one, which explains your results (I got that...

Hi ! For the scaling, the idea is try to scale such that all error terms (classification + position + size) have roughly the same scaling. Otherwise, the training would...

I am currently experimenting on how to fix the training. I set up a special branch `fix_training`. A few things I have noticed until now: * use a very simple...

@ithink2 Thanks for the testing. I am working on fixing this training problem, aiming to get at least ~0.7 mAP starting from the VGG weights. Things are getting a bit...

I had a quick check yesterday: this piece of code runs fine and returns the proper result on recent Nvidia GPUs (at least ML ones): ```python val_f16 = np.finfo(np.float16).smallest_subnormal val_f32...