Unstructured Pruning Demo

[karanchahal]

I was experimenting with pruning of object detection algorithms and was able to get some good results using the detectors in this repo. I though I'd share a popular technique for pruning via a demo here.

The algorithm used for pruning is from the "To Prune or not to Prune" paper that is incidentally used by the Tensorflow Model Optimisation Toolkit to perform pruning.

I was able to bring about a sparsity of 70% with almost no loss in mAP using this technique. I suspect getting upto a 90% sparsity is also possible with more training. Here I have only applied pruning masks to the convolutional filters (more specifically around 100 conv layers were pruned).

This branch of research is particularly exciting as one can (theoretically) get a lot of speed ups and compress the network more effectively. As pytorch is gearing up to support pruning and quantisation, I thought this would be a nice proof of concept. Maybe we can add support for general pruning into this repo by implementing the different algorithms in a subfolder.

Hope this demo isn't too out of scope for this repo.

Regards,
Karan

[karanchahal]

Google colab demo link

[fmassa]

Nice, thanks!

[Darshan2701]

@karanchahal Thanks for an awesome work.
Quick question - Can this be used to prune just the Faster RCNN model? Using resnet50-FPN as the backbone?

[karanchahal]

@Darshan2701 yes it definitely can. The good thing about this algorithm is that it's pretty general. Anything with a conv layer is prunable. Let me know it works out okay :)

[Darshan2701]

@karanchahal Thanks for the information, Karan. I will try to implement this and will let you know.

[92ypli]

@karanchahal Thanks for an awesome work.
Quick question - Can this be used to prune just the MobilenetV2 model?

[karanchahal]

Yes it should work. Try it let me know !

…

On Wed, Jul 24, 2019, 21:14 92ypli ***@***.***> wrote: @karanchahal <https://github.com/karanchahal> Thanks for an awesome work. Quick question - Can this be used to prune just the MobilenetV2 model? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#964?email_source=notifications&email_token=ADEXT7R3JK5LRJZ55OUSGJ3QBB2GJA5CNFSM4ICCBHH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2WYEVA#issuecomment-514687572>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADEXT7UBVMEJQVUK3IIAK53QBB2GJANCNFSM4ICCBHHQ> .

[enoceanwei]

Google colab demo link

Dear Karanchahal,

First of all, I am very interesting your jobs, it is an awesome work.

I have followed the codes and try to run it in my local machine.,However, my machine can not resolute the below code:

"
BBoxAugCollator()
"

In other words, I can not find the "BBoxAugCollator" API in my env libraries, could you tell me how to add the API or solve the problem.

my envs is : win10 x64, conda python 3.5, cuda10.0, cudnn7.1, pytorch=1.1.0 torchvision=0.2.1

I am looking forward to your reply, many thanks.

Regards

Wei

[NikolasBielski]

@karanchahal have you used this for Matterport's MaskRCNN?