Nevermore

Some results are shown in the Table. 5 in our [paper](https://arxiv.org/pdf/2103.14475.pdf).

Hi, these two implementions can achieve similar results. The fc-bn-swish-avgpool can be a little bit better (~0.1% top-1).

Not yet.

First, the Teacher netowrk is trained based on the dataset, and then we will freeze the Teacher's weight, finally the random initialized student will be trained by the GT and...

Hi, thanks for you attention to our work. Take the detection framework [mmdetection](https://github.com/open-mmlab/mmdetection) as an example, you can add our cmt network to the [backbones](https://github.com/open-mmlab/mmdetection/tree/master/mmdet/models/backbones), similar to ```pvt.py```, and use...

@lijoe123 You need to rewrite ```cmt.py``` and add it to ```/mmdet/models/backbones/```. As for the demo, you can refer to the [pvt](https://github.com/whai362/PVT/blob/v2/classification/pvt.py) for classification and [pvt](https://github.com/whai362/PVT/blob/v2/detection/pvt.py) for detection.

same problem. I find that some specific model (# parameters) with some specific batch size will encounter the error (```ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 40349)```), change the batch size...

We will release the corresponding vanillanet based Mask RCNN code and ckpts in several days.

Thank you for your interest in our work. For Q1 & Q3, we have only experimented with batch sizes of 1024 and 2048, and their results were similar. As the...

We will release the code and ckpt in several days.