search

zhaoweicai / mscnn

Caffe implementation of our multi-scale object detection framework
404 stars 211 forks source link

"detection_accuracy_layer" dosen't have gpu version #95

Open ubenz55555 opened 6 years ago

ubenz55555 commented 6 years ago

Hi zhaoweicai

Is your accuracy loss layer called "detection_accuracy_layer.cpp"

and your bounding box loss called "detection_loss_layer.cu" under mscnn/src/caffe/layers/ ?

I have noticed that your detection_accuracy_layer only have cpp version， which is

run by cpu only，and there is no detection_accuracy_layer.cu for gpu version ，but your

detection_loss_layer have both cpu and gpu version . Since you both use "detection_accuracy_layer.cpp"

and "detection_loss_layer.cu" to train our model， I wonder that you only use cpu to train your

model since there is no "detection_accuracy_layer.cu" . Is that right ? Thanks ~

zhaoweicai commented 6 years ago

Under the GPU mode, when a layer only has a CPU version, CPU will be used for its computation. But the computation for the other layers will still use GPU. Since the "detection_accuracy_layer" doesn't have too much computation, it is OK to use CPU for its computation.