TensorRT Plugin for implementing custom Eltwise caffe layer with 3 inputs


Is it possible to implement custom Eltwise caffe layer with 3 input tensors? I have gone through the sample plugins but still couldn’t make out how to build a custom layer. It is said in the documentation that the native Eltwise layer in TensorRT supports only 2 input tensors, does that mean TensorRT will not support 3 input tensors in Eltwise layer even if we go for custom layer implementation?

Any examples or sample codes in this regard would be of great help. Thanks in advance.


Which Eltwise operation do you use?

Suppose you can separate the 3 input case into 2x 2 input tasks directly?.
For example,

sum(a1,a2,a3) -> sum(a1, sum(a2,a3))


Hi AastaLLL,

Thanks for the suggestion. I’m using Eltwise Summation operation. Will try out your suggestion and see if it works.


Hi sajinmzac4,

Have you tried with our suggestions? Any result can be shared?


Hi kayccc,

The suggestion worked and it helped in resolving our error. Thanks a bunch. Would have been good if TensorRT would have supported the dynamic input feature as well. Looking forward to that in your next release.