Check failed: error == cudaSuccess (2 vs. 0) out of memory 错误。

When I use caffe for training recently, I often get   Check failed: error == cudaSuccess (2 vs. 0) out of memory   error.

It can be seen from the error message that there is not enough memory. So the editor added an 8G memory stick (just had it on hand), but it didn't work much, and still reported an error. A solution was found later.

Solution:

The batch_size is too large, and there are too many pictures read at one time, so it exceeds the video memory. Therefore, the batch_size of the files train and test in train.prototxt needs to be adjusted smaller.