alldbi / supermix Goto Github PK

Hi,

In the original paper of CutMix, it reports top-1 accuracy as 78.6 (21.4 top-1 error) on the ImageNet, however, in the Table.1 of your SuperMix, the corresponding result of CutMix is only 77.2. Could I know what makes the difference? Thanks.

HI

Thanks for your great work. I hope to use your supermix technology in object detection framework like yolov4/v5.
How can i integrated it efficiently in this framework?

Hi, @alldbi thank you for your interesting work

I'm trying to reproduce the results of table 1. on your paper, but I can't find the training code for it.
Or is it just using the train_teacher.py code with loading the produced SuperMix dataset? (+ by using get_cifar100_dataloaders_sample and class CIFAR100InstanceSample?)

When I was looking at the code, I suddenly noticed that you set the tol parameter to 70 percent by default, which means that if the batch is 8, only 2-3 supermix images will be returned. Is it because the performance of supermix is very good, so it does not need a lot of data？ in order to speed up the training, we set high tol parameters.

If supermix is used for image segmentation, should we replace KL with an IOU that evaluates x_hat with k inputs and then maximize it？

Hi,

In the original paper of CutMix, it reports top-1 accuracy as 78.6 (21.4 top-1 error), however, in the Table.1 of your SuperMix, the corresponding result is only 77.2. Could I know what makes the difference? Thanks.