dilation layer could not work in TensorRT 3.0 RC of Jetson TX2

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

This is my test prototxt file:
name:“denseASPP”
layer {
name: “data”
type: “Input”
top: “Input1”
input_param {
shape {
dim: 1
dim: 3
dim: 512
dim: 512
}
}
}
layer {
name: “Conv1”
type: “Convolution”
bottom: “Input1”
top: “Conv1”
convolution_param{
num_output: 64
pad: 6
dilation: 6
kernel_size: 3
}
}

when i ran command ./giexec --deploy=dilation_test.prototxt --output=Conv1 it can work
but when I ran ./giexec --deploy=dilation_test.prototxt --output=Conv1 --half2=true, it cannot work

2

Hi,

Dilated convolutions are not supported by TensorRT 3.
You can find this information in our document:

> 2.3.1.3. NvCaffeParser
...
Caffe features not currently supported by TensorRT include:
‣ Deconvolution groups
<b>‣ Dilated convolutions</b>
‣ parametric rectified linear unit (PReLU)
...

Thanks.

3

I have read the TensorRT 3.0 developer guide, it said that
NVCaffefeatures not currently supported by TensorRT include:
‣parametric rectified linear unit (PReLU)
‣leaky rectified linear unit (Leaky ReLU)
‣Scale (other than per-channel scaling)
‣ElementWise (Eltwise) with more than two inputs
Not exactly same as you posted. Could you please tell the link of your document?

Also, I have done the experiment, when I change the pad and dilation to 2, not 6, it worked.

4

I have read the TensorRT 3.0 developer guide, it said that
NVCaffefeatures not currently supported by TensorRT include:
‣parametric rectified linear unit (PReLU)
‣leaky rectified linear unit (Leaky ReLU)
‣Scale (other than per-channel scaling)
‣ElementWise (Eltwise) with more than two inputs
Not exactly same as you posted. Could you please tell the link of your document?

Also, I have done the experiment, when I change the pad and dilation to 2, not 6, it worked

5

Hi,

May I know your TensorRT version?
Do you install it via JetPack3.2?

Thanks

6

I install it via the leatest version of Jetpack 3.2. TensorRT’s version is TensorRT 3.0 RC2

7

I install it via the leatest version of Jetpack 3.2. TensorRT’s version is TensorRT 3.0 RC2

8

Hi,

We have doubled checked the document in JetPack3.2 DP, Dilated convolutions are not supported.

In /usr/share/doc/tensorrt/TensorRT-3-User-Guide.pdf.gz:

2.3.1.3. NvCaffeParser
...
Caffe features not currently supported by TensorRT include:
* Deconvolution groups
<b>* Dilated convolutions</b>
* parametric rectified linear unit (PReLU)
...

Thanks.

9

Is dilated convolutions supported on any other platforms?

10

Hi,

It is not supported by current TensorRT.
Please check if other frameworks have an API for it.

Thanks.