MATLAB Answers

int8 Calibration data with semantic segmentation

3 views (last 30 days)
Markus Walser
Markus Walser on 28 Oct 2019
Edited: Markus Walser on 28 Feb 2020
Hi
I successfully depoyed a fp16 network with GPU Coder. In order to accelerate speed I would like to test int8 resolution with the GPU Coder.
Therefore I changed the following option in the GPU Coder:
cfg.DeepLearningConfig.DataType = 'int8';
In the help is described that an additional data set must be provided to calibrate the network with limited resolution:
cfg.DeepLearningConfig.DataPath = 'C:\Temp\Matlab\Calibration\';
Even the documentation is not very clear how the provide calibration data for semantic segmentation I simply specified the logos_dataset from the TensorRT example for a first try. However as soon as I switch from fp16 to int8 I get the following error from GPU Coder:
??? Error generating Makefile while generating code for network 2019b-huc0102_0. TensorRT INT8 supports only classification networks. Output layer of network must be a
ClassificationOutputLayer
Are there any plans to support this semantic segmentation with int8 resolution in the future?
Best regards, Markus

  0 Comments

Sign in to comment.

Accepted Answer

Jaya Shankar
Jaya Shankar on 28 Oct 2019
Hi Markus
GPU Coder does not support INT8 inference for regression and semantic segmentation networks today. We will be looking to add this in a later release.
Jaya

  0 Comments

Sign in to comment.

More Answers (0)

Products


Release

R2019b

Translated by