TensorRT does not support in-place operations on input tensors in a prototxt file.

when i use caffeparser load model like this:
const IBlobNameToTensor* blobNameToTensor = parser->parse(“xxx.prototxt”, “xxx.caffemodel”, *network, DataType::kFLOAT);

it report :TensorRT does not support in-place operations on input tensors in a prototxt file.

the xxx.prototxt data Layer like :
layer {
name: “data”
type: “Input”
top:“data”
input_param{
shape{
dim:1
dim:3
dim:248
dim:248
}
}
}

if the xxxx.prototxt data Layer like this :
input: “data”
input_shape {
dim: 1
dim: 1
dim: 28
dim: 28
}

it’s ok,

i guess the data layer only has top no bottom,but i don’t konw how to solve it?(on x86 machine not Jetson TX1)

Hi,

A TensorRT network definition has no notion of in-place operation.

For example, the input and output tensors of a ReLU are different.
When a network uses an in-place operation, meaning that input and output have identical blob name, the TensorRT tensor returned in the dictionary corresponds to the last write to that blob.

It’s recommended to use different blob name for input and output blob to avoid non-necessary error.
Thanks.