No result when using tensorRT Sample FasterRCNN with other images

AastaLLL · October 11, 2017, 2:45am

Hi,

It should be DimsCHW(5,1,1).

Thanks.

r02525028 · October 11, 2017, 3:01am

Hi,

ok, thx

and if in caffe: top[0]->Reshape( bottom[1]->num(), 5, 5, 5 );
but in tesorRT we cannot return DimsNCHW(inputs[1].d[0], 5, 5, 5);

how to solve it?

AastaLLL · October 12, 2017, 5:45am

Hi,

There is something wrong in comment #21. I have corrected it.

For

shape {
dim: 300
dim: 5
}

This will automatically transfer to

shape {
dim: 300
dim: 5
dim: 1
dim: 1
}

That is:

BATCH_SIZE=300
DimsCHW = (5,1,1)

Thanks and sorry for the misleading.

OnePieceOfDeepLearning · October 12, 2017, 7:38am

one more question about im_info

in im_info, do we need to put the original size (height, width) or the resized size?

if I put resized size in im_info, the result is not what I expect.

r02525028 · October 13, 2017, 1:46am

Hi,

I face wrong result when doing convolution with dilation=2 (TensorRT2 on TX2)

and that may be the issue in

https://devtalk.nvidia.com/default/topic/1004959/?comment=5131210

my cudnn version is 6.0.21, is the same problem?
and if I update cudnn, should I do any setting for TensorRT2 on TX2?

AastaLLL · October 13, 2017, 3:25am

Hi,

The topic you posted is for Kepler GPU. TX2 is Pascal architecture, doesn’t have this issue.

Not sure what error do you face.
Do you have different accuracy between DIGITs and TensorRT? Or Caffe and TensorRT?

A most common issue in accuracy drop is the different mean subtraction implementation.
Please read this topic for more detail:
[url]https://devtalk.nvidia.com/default/topic/993552/jetson-tx1/detection-result-difference-between-jetson-inference2-3-and-digits5-1/post/5097211/#5097211[/url]

Thanks.

r02525028 · October 13, 2017, 5:15am

Hi,

in the layer (attached below) “res5a” (Eltwise), two input dimensions are different

and get message~ {res5a: all elementwise inputs must have same dimensions}

but in caffe, there is no such issue

and after test, I find input[1] of “res5a” with wrong dimension after “res5a_branch2b”

=========================================

input: “data”
input_shape {
dim: 1
dim: 3
dim: 224
dim: 224
}

layer {
bottom: “data”
top: “res5a_branch1”
name: “res5a_branch1”
type: “Convolution”
convolution_param {
num_output: 2048
kernel_size: 1
pad: 0
stride: 1
bias_term: false
}
param {
lr_mult: 1.0
}
}

layer {
bottom: “res5a_branch1”
top: “res5a_branch1”
name: “bn5a_branch1”
type: “BatchNorm”
batch_norm_param {
use_global_stats: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
bottom: “res5a_branch1”
top: “res5a_branch1”
name: “scale5a_branch1”
type: “Scale”
scale_param {
bias_term: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
bottom: “data”
top: “res5a_branch2a”
name: “res5a_branch2a”
type: “Convolution”
convolution_param {
num_output: 512
kernel_size: 1
pad: 0
stride: 1
bias_term: false
}
param {
lr_mult: 1.0
}
}

layer {
bottom: “res5a_branch2a”
top: “res5a_branch2a”
name: “bn5a_branch2a”
type: “BatchNorm”
batch_norm_param {
use_global_stats: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
bottom: “res5a_branch2a”
top: “res5a_branch2a”
name: “scale5a_branch2a”
type: “Scale”
scale_param {
bias_term: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
top: “res5a_branch2a”
bottom: “res5a_branch2a”
name: “res5a_branch2a_relu”
type: “ReLU”
}

layer {
bottom: “res5a_branch2a”
top: “res5a_branch2b”
name: “res5a_branch2b”
type: “Convolution”
convolution_param {
num_output: 512
kernel_size: 3
dilation: 2
pad: 2
stride: 1
bias_term: false
}
param {
lr_mult: 1.0
}
}

layer {
bottom: “res5a_branch2b”
top: “res5a_branch2b”
name: “bn5a_branch2b”
type: “BatchNorm”
batch_norm_param {
use_global_stats: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
bottom: “res5a_branch2b”
top: “res5a_branch2b”
name: “scale5a_branch2b”
type: “Scale”
scale_param {
bias_term: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
top: “res5a_branch2b”
bottom: “res5a_branch2b”
name: “res5a_branch2b_relu”
type: “ReLU”
}

layer {
bottom: “res5a_branch2b”
top: “res5a_branch2c”
name: “res5a_branch2c”
type: “Convolution”
convolution_param {
num_output: 2048
kernel_size: 1
pad: 0
stride: 1
bias_term: false
}
param {
lr_mult: 1.0
}
}

layer {
bottom: “res5a_branch2c”
top: “res5a_branch2c”
name: “bn5a_branch2c”
type: “BatchNorm”
batch_norm_param {
use_global_stats: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
bottom: “res5a_branch2c”
top: “res5a_branch2c”
name: “scale5a_branch2c”
type: “Scale”
scale_param {
bias_term: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

layer {
bottom: “res5a_branch1”
bottom: “res5a_branch2c”
top: “res5a”
name: “res5a”
type: “Eltwise”
}

layer {
bottom: “res5a”
top: “res5a”
name: “res5a_relu”
type: “ReLU”
}

layer {
bottom: “res5a”
top: “res5b_branch2a”
name: “res5b_branch2a”
type: “Convolution”
convolution_param {
num_output: 512
kernel_size: 1
pad: 0
stride: 1
bias_term: false
}
param {
lr_mult: 1.0
}
}

layer {
bottom: “res5b_branch2a”
top: “res5b_branch2a”
name: “bn5b_branch2a”
type: “BatchNorm”
batch_norm_param {
use_global_stats: true
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
param {
lr_mult: 0.0
decay_mult: 0.0
}
}

======================================

r02525028 · October 13, 2017, 9:08am

Oh,

I find that maybe the reason TensorRT2 doesn’t support dilated convolution

and it produce wrong result because pad size~ after updating to TensorRT3 it solved. Thx!!

by the way, if reshape

shape {
dim: -1
dim: 12
}

we should return DimsCHW = (12,1,1) right?

AastaLLL · October 16, 2017, 5:56am

Hi,

batch size of TensorRT is set on following code:

Set maximum batch size when creating the engine
[url]https://github.com/dusty-nv/jetson-inference/blob/master/tensorNet.cpp#L158[/url]
Set the real inference batch size when executing.
[url]https://github.com/dusty-nv/jetson-inference/blob/master/imageNet.cpp#L306[/url]

Beside BatchSize information, DimsCHW = (12,1,1).

r02525028 · October 16, 2017, 9:18am

Hi,

thx for your note!!

but if with two input

and we wanna get bottom[1]->num

how can we do that in tensorRT enqueue??

because batchSize represents for bottom[0]->num ??

AastaLLL · October 17, 2017, 8:03am

Hi,

BatchSize is identical across the networks.

r02525028 · October 24, 2017, 3:08am

Thx for your reply.

r02525028 · November 1, 2017, 1:59am

Hi,

when I test sample [FasterRCNN]

I revise the load image part to read .jpg

====================================
cv::Mat image = cv::imread( argv[1], CV_LOAD_IMAGE_UNCHANGED );

image -= cv::Scalar(pixelMean[0], pixelMean[1], pixelMean[2]);

unsigned volChl = INPUT_H*INPUT_W;

float* data = new float[NINPUT_CINPUT_H*INPUT_W];

cv::Mat_cv::Vec3f::iterator it, itend;

for (int c = 0; c < INPUT_C; ++c)
{
cv::Mat_cv::Vec3f::iterator it = image.begincv::Vec3f();
cv::Mat_cv::Vec3f::iterator itend = image.endcv::Vec3f();
for (unsigned j = 0; j < volChl; ++j){
data[(2-c)volChl + j] = float ((it)[c]);
it++;
}
}

and test this jpg image

,yet there are no correct result.

Note: the github jetson-inference using QT instead of OpenCV to load image

Is any suggestion for this problem?

AastaLLL · November 1, 2017, 3:04am

Hi,

QT format is RGBA while OpenCV is BGR.
Please remember to convert image format from CHW to HWC.

Thanks.

r02525028 · November 1, 2017, 7:38am

Hi,

I finally find we should resize image to match the input dims → 500*375 set in protofile

and it still using openCV to load image

then it works!

thx~

r02525028 · November 6, 2017, 1:43am

Hi,

if the layer is custom layer

how could we get the parameter in the protofile?

or we only have to key in them in our code?

AastaLLL · November 6, 2017, 3:03am

Hi,

IPlugin layer doesn’t support parameter parsing.
Please handle it with a self-defined file read/write.

Thanks and sorry for the inconvenience.

zheng_x_x · November 20, 2017, 1:16am