Closed apple2373 closed 5 years ago
I also guess that PFN internally have multi batch and multi gpu training of faster r-cnn as far as I read in PFDet paper (https://arxiv.org/pdf/1809.00778.pdf). Maybe they are not the chainercv team, but I wonder if you could take source code from them.
Hi. We currently have no plan of supporting those features for Faster R-CNN. However, we are working on FPN, which supports those features. https://github.com/chainer/chainercv/pull/685
The same feature request is raised in StackOverflow.
Faster-RCNN with FPN is supported for multi batch GPU. https://github.com/chainer/chainercv/blob/master/examples/fpn/train_multi.py For normal Faster-RCNN, we are planning to rewrite it and support multi GPU training.
Thanks for the multi gpu script, but I can't install chainermn so I couldn't use it. Can you also show the script without chainermn?
Hi,
I'd like to request to support multi gpu and multi batch training for faster R-CNN. I think these issues are not recognized by the developers because I couldn't find the issue for them, so I am creating the issue.
(Well, sorry if you are already working on them. I tried to search from the past issues, but I am not so confident, because GitHub search function is not so good.... I referenced the most relevant issue that I was able to find. )
1) Support multi batch training It seems to support only batch size = 1. I found a similar issue (#589) but it's just closed saying that it only support batch size of 1, which is not good! Please support it!
2) Support multi GPU training Not only batch it should support multi GPU training. I found a same issue for YOLO (#627) but couldn't find Faster R-CNN one, so I am requesting here.
Lastly, this is just FYI, but I am requesting these functions because they are already supported in popular pytorch implementations.
https://github.com/jwyang/faster-rcnn.pytorch https://github.com/facebookresearch/maskrcnn-benchmark
Thanks!