Relabel data about relabel_imagenet HOT 6 CLOSED

@123wk45678
The code for computing the prediction map is not provided yet and we may update it later.
Getting ReLabel data for your own data is not that hard. You can use a pre-trained strong classifier (e.g., EfficientNet-L2) and get the prediction map by removing the global average pooling layer as described in our paper's appendix.

from relabel_imagenet.

ok,thank you ,expect this to open source your code ,author

from relabel_imagenet.

A bit confusing here.As described in paper,remove the global average pooling layer and add conv1x1 layer, do we need to fine-tuning this conv1x1 layer for getting label map?

from relabel_imagenet.

@changewOw
Indeed, the 1x1conv layer is exactly the same as the final fully-connected layer.
The size of the fully-connected layer's weight matrix [d x C] (d=channel size, C=number of classes),
and we reshape it to [1 x 1 x d x C] to be the kernel matrix of 1x1conv.
We present the detailed descriptions in Appendix A of our paper.

from relabel_imagenet.

Got it! Thank you very much

from relabel_imagenet.

@changewOw You're welcome :)

from relabel_imagenet.