larspars
commented
4 years ago Hi! We did a little dabbling on ImageNet, but not a lot. On my hardware (somewhat old GPUs, slow disks) training on ImageNet takes a very long time, so it's hard to iterate meaningfully. We were able to get 43.4% top-1 error on the validation set, which is close to AlexNet, but my suspicion is that this may be far from the ceiling. This was using 112x112 images, and splitting up the model over two GPUs (first layers on one GPU, last layers on the other). I don't think I have the code for this in a usable state unfortunately. For the paper we wanted to keep the convnet architecture relatively uniform across all experiments, but I suspect better results could be seen with a different architecture. The activation vectors really blow up in size when using larger images, so I suspect there's room for improvement in handling that in a different way.
jlindsey15
Hello! Thanks for making this code available. I wanted to ask what your experiences / results have been running this algorithm on ImageNet, and how this code may be used to do so (when I try setting ImageNet as the dataset, my GPU runs out of memory -- I can imagine various ways to address this, but I wanted to see if you have encountered a similar issue and if so, how you've addressed it). Thank you!