TensorRT 2.1.2 sampleINT8 with googleNet

Accelerated Computing GPU-Accelerated Libraries
1

I have succesfully installed TensorR 2.1.2 and execute the samples, including sample_googlenet and sample_int8 with the MNIST dataset.

I tried to execute also the sample_int8 with the googlenet network but it is throwing a SEGMENTATION FAULT. I guess it is due to the lack of a batches folder that can be used for calibration.

How could I provide that folder or diagnose the problem?

Context:

CUDA 8.0.26
cuDNN 6.0.21
tensorRT 2.1.2
driver nvidia-375

2

The documentation here provides some instructions on generating batch files:

3

Yes, I read the documentation. But it is still unclear for me how to proceed in the following situation:

  • I have a pretrained network defined in Caffe (.prototxt, .caffemodel).
  • I have a collection of representative images. The only preprocessing needed is normalize against max value.
  • I have Caffe already installed.
  • When running inference with pyCaffe, I only need to load the images with imread, and pass them to the network without any other pre-processing than adjusting the batch and channel position (values are loaded in float32, between 0 and 1).