search

keyu-tian / SparK

[ICLR'23 Spotlight🔥] The first successful BERT/MAE-style pretraining on any convolutional network; Pytorch impl. of "Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling"
https://arxiv.org/abs/2301.03580
MIT License
1.45k stars 84 forks source link

Optimized sparse conv from open3d #30

Closed ifeherva closed 1 year ago

ifeherva commented 1 year ago

This repo is using masked dense convolutions because it is optimized in torch. However, would this implementation speed things up or too complicated getting this work here?

keyu-tian commented 1 year ago

Thanks for this! This framework looks easy to use but it doesn't support depthwise or grouped convolution, while many CNNs rely on them (resnext, convnext, mobilenet, etc.). We are still using masked conv for now since we found no framework currently has a sufficiently efficient implementation (i.e., faster than masked conv in torch) of sparse depthwise/grouped convolution on GPUs.

xubin04 commented 9 months ago

How about torch sparse++?