Created by: jczaja

Hi,

I want to solve following problem: Activations of paddle (Grad ops) are using inputs: Out, Out@GRAD while mkl-dnn grad primitives activations need: X (input signal to fwd op) . eg. MKL-DNN grad activation ops are expecting input of forward to be available. So far we used SetBlob to have it transfered, but this is problematic for parallel executor. Hence, please advice how can I modify paddle so that In activation grad ops (for mkl-dnn builds), X (input signal to fwd op) is available in grad op ? In other words : ExecutionContext::OperatorBase::inputs_ should have: {["Out"] = vector<> , ["Out@GRAD"] = vector<>, ["X"]= vector<>}

Please advice on how to do so if it is possible?