Caffe FCN Test | Check Failed:error = = cudasuccess (2 vs. 0) out of memory

When using Caffe run FCN test, the following error is prompted:
Check Failed:error = = cudasuccess (2 vs. 0) out of memory
is the memory overflow, and is the Cuda, then the GPU cache overflow.
In this case, the first thing to see is the training and testing of the batch_size, and the training process is no problem, indicating that the training batch_size no problem;
This problem with the test process, see the FCN of the source code in the section on testing:

Here's reshape (1,*), 1 of which means batch_size, so there is no way to adjust.
So also suspected is not the use of the original image is too large, occupy more space, and the revision still reported this error, so can only first not test, direct training.
I would like to discuss it with a small partner with the same problem.