Training result is too sparse

[xingyueye]

@ShuangXieIrene Hi, I use this proj to train my own dataset(a small and simple scene). I found that the trainning result is so sparse, it means lots channels of weights are zero, that it's hard to quantilize. I wonder if there are some Regulation operate in this proj that I ignored it. Could you help me with problem, thanks a lot

[foreverYoungGitHub]

hi @xingyueye, did you use the pretrained weights to do the fine-tuning or start the training from scratch?

[xingyueye]

hi @xingyueye, did you use the pretrained weights to do the fine-tuning or start the training from scratch?

I used both, the results are simliar. I guess if it is ReLU6 caused the sparsity.

[foreverYoungGitHub]

Might be, espectially for the small dataset. But previously when I train a detector for a tiny dataset(~400 Images), the weight of the networks is not that sparse. In that task, I used resnet18 with yolov3. Would you mind sharing which backbone and detection head are you used to train the network?

Btw, would you mind try to train the model with the ssds.pytorch in the dev branch. Cause I did not work with the code in the master branch for a long time, maynot able to provide too much advise for the old version code.

[xingyueye]

thanks, I would try the dev branch. I used to train mobilenetv1 + ssd_lite on a very small dataset(3 classes and the objects are very simple)

[foreverYoungGitHub]

Cause by the depthwise/pairwise conv, the mobilenetv1 + ssd_lite combination is easy to have the sparse weight. If the final deplyment enviroment is using nvidia gpu, I would suggest to try resnet or regnet and yolo for this simple experiment.

Let me know if you meet any problem when you use ssds.pytorch.