Created by: wojtuss

I'm working on fusing FC and GRU layers by the MKL-DNN GRU op for inference. Using the inference_transpiler.py, is it possible to grab the weights from the FC's MUL op and "add" it as a weight to the MKL-DNN-based GRU op?

The situation is as follows:

• in the GRUOpMaker there is a new input WeightX added (AsDispensable()),
• there is no WeightX input passed via dynamic_gru in nn.py, so the ordinary GRU could work,
• a new GRUMKLDNNKernel is created, which requires the WeightX tensor, it is called only when inferencing with MKL-DNN,
• in the inference_transpiler.py I am fusing the MUL+elementwise_add+GRU ops, but this requires grabbing the MUL's weights into the GRU's WeightX input. How to achieve that? From the code of the transpiler, I am not really sure how to do that.

I know the "Pass" is comming to facilitate inference fusing, but until then I want to familiarize myself to and do the fusing using the transpiler.