search

varunagrawal / tiny-faces-pytorch

Finding Tiny Faces in PyTorch
MIT License
164 stars 44 forks source link

Runtime efficiency improvement for computing distances between every two bounding boxes #21

Closed fengyuentau closed 2 years ago

fengyuentau commented 3 years ago

Thanks for your great work.

When I try to cluster the whole training set (159,424 bboxes), I found the speed for computing distances between every two bounding boxes is too slow, which requires ~40hrs to finish with a Intel(R) Xeon(R) Gold 6132 CPU @ 2.60GHz (56 cores).

So I rewrite the corresponding code so as to take the advantage of Numpy's parallelism. Now it can finish in 9mins.

However, the clustering itself is still pretty slow. I do not hack into the pyclustering though, I have no idea how much time it will need to finish with all the bboxes.

fengyuentau commented 3 years ago

Well, it turns out that doing clustering in Python with this pyclustering is a nightmare and takes forever to finish. I turned back to the official matlab code for clustering all bboxes. Sad that matlab code is too hard to understand for me.