INT8 dnn error using dwDNN_initializeTensorRTFromFile

Autonomous Vehicles DRIVE AGX Xavier DRIVE AGX Xavier General
1

Hi,

We are trying to infer a semantic dnn on AGX. The model was created with tensor flow.
The optimization tool tensorRT_optimization generate a bin file without mentioning any error.

However, when using the function dwDNN_initializeTensorRTFromFile we got the following error:
cudnnSerializationUtils.h:67: const T& nvinfer1::cudnn::extractBlobs(const ifb::LayerParams&) [with T = ifb::CaskDeconvolutionBlobs; ifb::BlobsUnion tag = (ifb::BlobsUnion)5u]: Assertion `x == tag’ failed.

FYI, the same model in FP32 and FP16, does not produce any errors…

What could be the problem?
Thanks in advance,
Sacha

2

Dear sacha,
Could you please details of network and UFF file to reproduce the error on ourside

3

Dear SivaRamaKrishna,

Could you please send a link where to upload the file.

Best,
Sacha

4

Dear Sacha,
You can provide google drive link. If not, Could you raise a bug and share Bug ID?

5

Dear SivaRamaKrishna,

On the link below you will find the uff file:
https://drive.google.com/file/d/16ZR16kuMRLh0SD3I1ljdS8HBrKE6wYC9/view?usp=sharing

Best,
Sacha

6

Hi Sacha,
Could you please share network details such as input Dimensions, input/output blob names which are required by tensorRT optimization tool.
Also, it would be great if you can share the type of layers in your network. We can check if any unsupported layers causing the problem.

7

Hi SivaRamaKrishna,

input dimension: 3x1920x1208
output dimension: 6x480x302
inputBlobs = features
outputBlobs = probabilities

8

Dear sacha,
I have used the following command and generated TensorRT model and able to load in without any issue on my Host PC.
/usr/local/driveworks-1.2/tools/dnn/tensorRT_optimization --modelType=uff --uffFile=/path/to/uff --int8 --inputDims=3x1208x1920 --inputBlobs=features --outputBlobs=probabilities.

Could you please confirm you are facing issue in dwDNN_initializeTensorRTFromFile call by keep exit(0) after it.

9

Dear SivaRamaKrishna,

the command you launched optimize in FP32, it must me “–int8=1” and “–calib=batch0”.
Here you have the calibration file:
https://drive.google.com/file/d/1qWfU-ZCNJ19N4iAuO2ccAT04L4MdQ1bO/view?usp=sharing

Best,
Sacha

10

Dear sacha,
Thank you for providing details. I am able to reproduce it on my end. Let me check with internal team and update you

11

Dear sacha,
Could you please check this symptom with our latest release. I don’t notice any error when using the function dwDNN_initializeTensorRTFromFile.
Please find our latest release(sdkmanager-0.9.9.2351) at https://developer.nvidia.com/nvidia-drive-downloads