No result when using tensorRT Sample FasterRCNN with other images

Autonomous Machines Jetson & Embedded Systems Jetson TX2
#1

Hi,

when I using the tensorRT FasterRCNN sample

with the image offered by github in dusty-nv/jetson-inference

ex: jetson-inference/dog_2.jpg at master · dusty-nv/jetson-inference · GitHub

and other images, there are no detection results even though I lower the threshold

“nms_threshold” and “score_threshold”. I have revised the code to load .jpg and

it works when loading images in sample faster-rcnn folder transformed to .jpg from .ppm

#2

Hi,

In FasterRCNN sample, we extract image data via function ‘readPPMFile’.
readPPMFile is a simple parser for the uncompressed image format. It can’t read a .jpg file, and the output will be incorrect.

For jpg decoder, you can check this sample.
[url]https://github.com/dusty-nv/jetson-inference/blob/master/imagenet-console/imagenet-console.cpp#L68[/url]

Thanks.

#3

Thx!

And another question,

can we adjust OutputDimension in stage enqueue?

like in caffe, we can reshape top blob in stage Forward_cpu after LayerSetup.

Am I right to seem getOutputDimensions() and initialize() as LayerSetup in caffe

and enqueue as Forward_cpu in caffe?

#4

Hi,

No.

When creating the TensorRT engine, the output dimension is chosen for optimized performance.
TensorRT doesn’t support the dimension change on the fly. Sorry for the inconvenience.

#5

ok!

I load trained model but it break:

[libprotobuf FATAL …/…/…/externals/protobuf/aarch64/10.0/include/google/protobuf/repeated_field.h:1078] CHECK failed: (index) < (current_size_):
terminate called after throwing an instance of ‘google_private::protobuf::FatalException’
what(): CHECK failed: (index) < (current_size_)

I have refered https://devtalk.nvidia.com/default/topic/1004468/how-to-start-tensorrt-on-tx1-/

but it still doesn’t work.

what kind of problem may it happened?

#6

Hi,

Which Caffe version do you use for training?
Please remember that TensorRT2.1 only support nvcaffe-0.15 branch.

NvCaffe-0.16 may lead to error.
Thanks.

#7

I train by using BVLC/caffe(version:1.0.0-rc3) instead of using DIGITS.

Does it affect?

Should I retrain my model by nvcaffe-0.15 instead?

#8

Hi,

Do you use TensorRT1.0?
Could you try to use TensorRT2.1 with JetPack3.1?

Thanks.

#9

No,

I use TensorRT2.1

#10

Hi,

Could you share prototxt and caffemodel?

#11

Hi,

I revise

“==========================”
input: “im_info”
input_shape {
dim: 1
dim: 3
}
“==========================”

to

“==========================”
input: “im_info”
input_shape {
dim: 1
dim: 1
dim: 1
dim: 3
}
“==========================”

in protofile, and it works, why it works after revised?

(note: it breaks in “Begin parsing model…” before revised)

#12

Hi,

They have different meaning:

input_shape {
dim: 1 ← BATCHSIZE
dim: 1 ← channel
dim: 1 ← height
dim: 3 ← widith
}

Please check this tutorial for more information:
http://caffe.berkeleyvision.org/tutorial/net_layer_blob.html

#13

Hi,

I face another problem

“================================================================”
Begin building engine…

res5a: all elementwise inputs must have same dimensions
res5b: all elementwise inputs must have same dimensions
res5c: all elementwise inputs must have same dimensions
“================================================================”

I remember tensorRT2 support elementwise, but why this message appear?

#14

Hi,

From the error message, your elementwise has different dimensions.
Elementwise need two identical dimension input. Could you recheck it?

#15

Ok, I will recheck it thx!

When implement getOutputDimensions(int index, const Dims* inputs, int nbInputDims)

what is nbInputDims exactly?

and inputs[0].d[0],inputs[0].d[1],inputs[0].d[2],inputs[0].d[3] are always (N,C,H,W)?

If not, how to get N of (N,C,H,W)?

#16

I have the same problem.

And this is due to the data shape defined in prototxt.

input_shape {
dim: 1
dim: 3
dim: 375
dim: 500
}

the dim has to fit the height and width of the input image.

in the original faster rcnn, I can pass any resolution by reshape input layer before forwarding.

Can I reshape the input shape in TensorRT?

#17

Hi,

It looks like this issue is from your model design.
You can check it via DIGITs or other visualized frameworks.

Dimension of C, H, W can get by getOutputDimensions(…)
[url]Face-Recognition/pluginImplement.cpp at master · AastaNV/Face-Recognition · GitHub

getNbOutputs() is return the number of output blobs, rather than the dimension.

Thanks.

#18

Hi,

so if in getOutputDimensions, we just return DimsNCHW

and in its next layer, we can get

inputs[0].d[0], inputs[0].d[1], inputs[0].d[2], inputs[0].d[3] for (N,C,H,W) right?

and DimsHW in its next layer, we can get

inputs[0].d[0], inputs[0].d[1] for (H,W) right?

#19

Hi,

Only C, H, W information is available.

In plugin API, user doesn’t need to take care of BatchSize handling.
The operation will be applied to all the batch data automatically.

Thanks.

#20

Hi,

====================
reshape_param {
shape {
dim: 300
dim: 5
}
}

so in such reshape param

we just return DimsCHW(300,5,1) or DimsCHW(1,300,5)??