Put some nodes of the same network on the GPU and others on the CPU?

Question

When defining a network in Caffe / Caffe2, can you host some nodes on the processor and others on the GPU? If so, how?

(If your answer is for a specific version of Caffe, indicate which one)

Answer 1

In general, the answer is NO : you cannot configure the device for each level yourself for the reasons Pooya Davoodi and Harsh Wardhan .

, , , . , , , GPU, ( , ).
"Python": , word2vec.
, , , .

---

, caffe2? PATENTS?!
UPDATE: , fb caffe2. !

Answer 2

, . solver.prototxt, , CPU GPU, . . , CNN, . CPU , . , , . , , CPU-GPU. , GPU CPU PCIe, , . , . , CNN, GPU, .

Answer 3

This may be possible in Caffe2, but I have never tested it. In Caffe2, each blob and operator has a device assigned to it. The operator is launched on the device assigned to him. But then you need to take care of initialization and communication manually, because the data_parallel_model in Caffe2 is only for installation with several GPUs.