tensorrt and caffe output have large difference with probability of a few percent

86108429 · April 8, 2019, 3:14am

Hello, I have a problem.

I test 100 images, and find that, for tensorrt and caffe output(such as softmax layer), although most images have very small difference, about 0.0x or 0.00x. but several images still have very large difference, about 0.x. This phenomenon occurs for tensorrt4 and tensorrt5.
So what causes this? can you give some advises?

86108429 · April 12, 2019, 8:28am

As the image preprocessing use opencv for tensorrt, so if the image preprocessing also use opencv for caffe inference, and the output will be same. Close this problem now.

Topic		Replies	Views
use int8 the result would be 0.1% differ TensorRT	3	643	June 19, 2019
Plugin layer is much slower than caffe TensorRT	0	553	June 11, 2018
Using tensorrt to transform Caffe's trained model TensorRT tensorrt	2	360	June 15, 2020
TensorRT softmax layer very slow GPU-Accelerated Libraries	3	1490	February 5, 2018
FCN Segmentation: Imported model with Tensorrt gives significantly worse results than pycaffe TensorRT	6	1223	October 12, 2021
accuracy drops when using FP32 in faster rcnn Jetson TX2	4	636	October 18, 2021
tensorrt inference result is wrong and inconsistent with caffe inference result TensorRT	2	915	February 27, 2020
TensortRT vs NVCaffe General	1	3589	October 16, 2017
TensorRT 5 batch size TensorRT	1	971	March 25, 2019
Get warning "Bias weights are not set yet" when converting caffe model with custom plugin TensorRT tensorrt	8	566	October 12, 2021

tensorrt and caffe output have large difference with probability of a few percent

Related topics