It is now possible to access data that is on the GPU via Python, thanks to pull request #5904 from Jon Long, which is based off of a patch by Takuya Narihira.
Previously, custom Python layers would only be able to transfer data via the CPU. While this is often fine for Python data layers, hidden Python layers would need copy the data back and forth between the CPU and GPU. Now, libraries such as PyCUDA can be used within custom layers without this needless “blob” re-routing.
PR python caffe1