In caffe’s definition of Eltwise layer, a coeff is introduced for sum operation (http://caffe.berkeleyvision.org/tutorial/layers/eltwise.html)
repeated float coeff = 2; // blob-wise coefficient for SUM operation
Does tensorrt’s caffe parser honor this coeff field?
Hello,
For a list of supported TRT ops, please reference:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#support_op
I found the coefficient is not honored when running my model in tensorrt. Could you confirm from source code that it is used? The docs you posted does not mention the coeff field.
I will verify with the engineering team. In the meantime, it’d help us debug if you can provide a small repro demonstrating how coeff field is ignored by trt.
You can just run a program that only contains one eltwise sum layer:
layer {
name: “input1”
type: “Input”
top: “input1”
input_param {
shape {
dim: 1
dim: 1
dim: 1
dim: 1
}
}
}
layer {
name: “input2”
type: “Input”
top: “input2”
input_param {
shape {
dim: 1
dim: 1
dim: 1
dim: 1
}
}
}
layer {
name: “mysum”
type: “Eltwise”
bottom: “input1”
bottom: “input2”
top: “result”
param {
lr_mult: 0
decay_mult: 0
}
eltwise_param {
operation: SUM
coeff: 0.5
coeff: 3
}
}
If your input is 1 and 1, you should get 3.5 instead of 2
hello,
just got confirmation TRT caffe parser does NOT support Eltwise layer’s coeff field.
Thanks