DS-Net for object detection

[NoLookDefense]

Hello. Thanks for your work. I noticed that you also conducted some experiments in object detection. I wonder whether or when you will release the code

[changlin31]

Hi, @NoLookDefense Thanks for your interest. Our experiments on object detection is conducted base on another Github repo (https://github.com/lzx1413/PytorchSSD). Please also refer to issue https://github.com/changlin31/DS-Net/issues/4

[NoLookDefense]

Thank you. I noticed that the code is recently updated two years ago. Is it just an SSD implementation, without your algorithms like slimmable NN and gating implementation?

[changlin31]

Yes, we currently don't have plans to release the detailed code for object detection. Sorry for any inconvenience.

[twmht]

@changlin31

Do we need to use the pretrained weight trained from slimmer algorithm if we want to train object detection?

[changlin31]

@changlin31

Do we need to use the pretrained weight trained from slimmer algorithm if we want to train object detection?

Hi @twmht

We used the dynamic slimmable pretrained weight for our object detection results in the paper. But I expect performing slimmable training on object detection task with a normally pretrained network is also possible. (We are getting good results with the second training approach on action recognition tasks)

[twmht]

@changlin31

What is your second training approach? Did you mention in the paper?

[changlin31]

@twmht

By second approach, I mean: first, directly load a normal ImageNet pre-trained model (e.g. pytorch ResNet checkpoint) and then, perform supernet training and gate training on the downstream tasks.

I did not mention this in the paper. This is a work we are currently working on.

[twmht]

@changlin31

did you also use in-place distilling with object detection? Which loss did you use for the classification branch?