Xueyan Zou

Actually, when I train on 2 GPUs and 4 GPUs machine. The performance did variate, with a 2 percent drop on 4 GPUs machine. From my point of view, as...

@VainF if my experiment did has any problem. Please point out!

please set -arch on nvcc cmd in make.sh file. like the following `nvcc -c -o roi_pooling.cu.o roi_pooling_kernel.cu \ -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC $CUDA_ARCH -arch sm_52` I highly recommend...

You should check your cuda version. There architecture is related to your cuda version instead of GPU version. Please refer to the link https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/

I also had 95% on resnet 18. This is extremely strange. This number is far different from what I could get from Kaiming's paper and also other person's paper. I...

Me too.....

I use ground truth flow for training. Generated Flow and Flow masks are only used in the evaluation. Thanks!

Thanks for your interest. E.g. The flow is computed based on Img1 -> Img2, if we want to warp back from Img2 to Img1 pixel by pixel. We need to...

Your training speed doesn't seem to be wrong, one trick I developed for the next project is to precompute the language encoder weight: https://github.com/UX-Decoder/FIND/blob/708ddf53ab594fe6be642bae2ff54eb42cdb8a9a/configs/grin/focalt_unicl_lang.yaml#L58. If you are interested in this,...

Use interactive segmentation as an example, here is the matching code: https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once/blob/7b2e76dbb17d0b7831c6813a921fe2bc8de22926/modeling/modules/matcher.py#L314, noted that as there are positive and negative examples so that masking is employed.