search

NVIDIA / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
672 stars 263 forks source link

I think there are two bugs in this version. #523

Closed ghost closed 6 years ago

ghost commented 6 years ago

Recenlty I train https://github.com/fwang91/residual-attention-network with this nv caffe version, but the loss stay around 8.0, and don't descrease. After a few debuging. now the loss can decrease, so I think there are two bugs.

the first is(eltwise,sum): if (op_ == EltwiseParameter_EltwiseOp_SUM && nocoeffs) { for (int i = 0; i < bottom.size(); ++i) { bottom[i]->ShareDiff(top[0]); } top[0]->ShareData(bottom[0]); } It should not be shared, if does, network such as resnet may not work.

and second the split layer,
bottom[0]->ShareDiff(*top[0]); diff should not shared for the same reason.

drnikolaev commented 6 years ago

@zhuwnq this is fixed in new release, coming soon...

drnikolaev commented 6 years ago

@zhuwnq could you verify https://github.com/drnikolaev/caffe/tree/caffe-0.17 release candidate?

drnikolaev commented 6 years ago

@zhuwnq Please verify https://github.com/NVIDIA/caffe/tree/v0.17.1 release and reopen the issue if needed.