Zhiqiang Shen

@lichengunc I think you're right. I also achieve higher performance after sorting with ppl.

Hi @pawopawo, this issue should be solved now. "shake_shake" has some unique components and hyper-parameter choices which are slightly different from conventional settings of VGGNet, ResNet, etc., so we use...

Hi all, sorry for the trouble aroused by our negligence. I will upload the pre-trained models soon.

Hi all, please check https://github.com/AaronHeee/MEAL/commit/2210139d708b31604d07e0c71bd5888af9dd6885.

@kaishijeng Here are the solver.prototxt and train_val_densenet.prototxt files for densenet_121. I think you can use them to train caffe models on ImageNet without any modification. @liu666666 @sunformoon Now I work...

Hi @Xglbrilliant, thanks for your interest in our work! Our FKD-SSL is based on our previous work: [S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-bit Neural Networks via Guided...

Hi @qijiezhao, the weights are not shared. In our structure, different scale features have different gates to re-scale the activations. So each resolution has an independent gate module.

Hi @zjuzuhe, thanks for pointing out this, actually, we did not use relu following the scale operation in global-level attention. I will remove or comment this line soon. I'm also...

@DL-ljw, batch_size can be a small value (6 is fine), but accum_batch_size should be 128 or larger.

Hi @DL-ljw, there is no voc2012 annotation available. So you should test voc2012 testset separately and submit the results to the evaluation server. You only need to change line 29...