something wrong with the test of SE-ResNet-50? about senet HOT 3 CLOSED

The validation images have not been changed since ILSVRC12, but the order of labels has been changed since ILSVRC13. You can take the order-changed validation list here. In addition, the mean values we used are [104, 117, 123] rather than decimals, which we listed at the beginning of each model prototxt.

And if you want to achieve exact results, I suggest you using ImageDataLayer, resizing the short edge of input images to 256 (To convenient implementation , you can do a small modification to function ReadImageToCVMat() in src/caffe/util/io.cpp ) and than coducting center crop 224x224 as described in the caption of Table 1.

Looking forward your testing feedback.

from senet.

Thanks so much. Following your suggestions, I replace the labels, and get an top-1 accuracy of 76.96%, and top-5 accuracy of 93.42%. It's really a great improve compared with the baseline, but a little lower than the proposed result. (I have resize the images to 256x256) Is there any other tricks when doing the testing?

Thanks again for your suggestions and such excellent work.

from senet.

There are no other tricks for testing. And I think you can achieve the proposed results after resizing the input images to 256 x N, which is a general approach for testing.

from senet.