A bug for train resnet152_trident.cfg using sparsity approach

[May-forever]

Dear @AlexeyAB,

I am so sorry to bother you, but could you please do me a favor?

Windows10+Pytorch 1.1.0

I find a bug when using the Slim-yolov3 repo: https://github.com/erikguo/yolov3

When I sparsity train resnet152_trident.cfg on my custom dataset. an error occur when I input the

following command:

python train.py --data data/obj.data --cfg cfg/resnet152_trident.cfg

I find that all the Resnet-structure can not been sparsity train on this project.

Do you know what should I do now? I would be appreciated if you can give me a hand.

Looking forward to hearing from you.

Thank you very much.

Best regards, May

[AlexeyAB]

Try to ask @erikguo or @glenn-jocher to support TridentNet if it is necessary in their project. More: https://github.com/AlexeyAB/darknet/issues/3363

[May-forever]

Try to ask @erikguo or @glenn-jocher to support TridentNet if it is necessary in their project. More: #3363

Thank you very much for your reply.

I think sparsity training and are important as the final weight are smaller than the original weight.

This is great for mobile deployment.

Will model pruning be integrated on darknet ?

[AlexeyAB]

@May-forever

May be you just can reduce number of filters 2x and train from the begnining without pre-trained weights, and you will get ~ the same effect: https://github.com/AlexeyAB/darknet/issues/3732

[May-forever]

@May-forever

May be you just can reduce number of filters 2x and train from the begnining without pre-trained weights, and you will get ~ the same effect: #3732

Dear @AlexeyAB ,

Thank you very much for your reply.

I have tried to set all filters=filters/2.

However, there is no theoretical support for using the method by just set filters=filters/2 to train a

network, and it is quite hard to write a new paper.

I would be appreciated if the model pruning can be integrated in Darknet.