Li wen

yes, you are right. But a very interesting thing is the wrong initialization doesn't harm the result, though the loss is large at the beginning. This is a finding in...

CUDA 8.0 failed too But if I use *openblas* instead of *mkl*, complie correctly but fail in running. and I modified the *.cu* file where $NewSteam()$ to $NewSteam(0)$, or there...

beside, It seems that cuda 9.0 doesn't support gcc-7

I think it is not a good idea to use this box free strategy in complex object detection. This scale regression strategy works in simple object detection tasks like pedestrians...

I have run the original code, and get 12 MR. I also implement the code in PyTorch, and get 13 MR. Both batch size is 8. (4 x 2GPUs). @hdjsjyl...

nvidia P40 has 24GB mem. And in pytorch, 4 images per batch only cover about 9~10GB mem. I don't like keras by the way. It hides too many details.

> > nvidia P40 has 24GB mem. > > And in pytorch, 4 images per batch only cover about 9~10GB mem. > > I don't like keras by the way....

The codes may contain some mistakes, currently just me is working on it. I hope other people could give some advice to reproduce the result.

Actually there is one thing confusing me so much. usually in object detection, we will freeze the batchnorm layers and the first block in ResNet if pretrained weight is used....

> @lw396285v Because CSP use a large batchsize, updating BN can improve the performance in SNIPER: Efficient Multi-Scale Training. Thanks for the reply. But you may make some mistakes, I...