TensorRT 2.0.4 aborted when testing ResNet50 using int8

Accelerated Computing CUDA CUDA Programming and Performance
1

Testing the ResNet50: deep-residual-networks/ResNet-50-deploy.prototxt at master · KaimingHe/deep-residual-networks · GitHub with giexec from gie_samples returns a Aborted:
./giexec --deploy=resnet50.prototxt --output=prob --int8
deploy: resnet50.prototxt
output: prob
int8
Input “data”: 3x224x224
Output “prob”: 1000x1x1
giexec: cudnnBuilder2.cpp:1157: void nvinfer1::builder::addReformatLayers(nvinfer1::builder::Graph&, nvinfer1::builder::Format::Type, std::unordered_map<std::basic_string, float>*): Assertion `regionScales->find(t->region->name) != regionScales->end()’ failed.
Aborted (core dumped)

Testing without --int8 works well.
Testing with googlenet.prototxt from gie_samples and with --int8 works well too.

Could anyone tell me what’s wrong with it? Thanks.

Hardware: TITAN X PASCAL, E5-2630V3
Software: CUDA 8.0.61

2

Hi，

I meet the same problem. I using TensorRT in my project and get the same wrong information.
Did you solve this problem?

Thanks

3

I have no idea about this problem, but INT8 inference provided by cuDNNv6 works well.

4

Hello Jinwei,

I am using Nvidia 1080 ti GPU. I have trained MNIST model in FP32, I would like to know how did you make your setup with RESNET to inference using TENSORRT. Can you provide your setup details and can you also let me know how TENSORRT does FP32 to INT8 optimization ?

Thanks in advance.