search

ceccocats / tkDNN

Deep neural network library and toolkit to do high performace inference on NVIDIA jetson platforms
GNU General Public License v2.0
718 stars 209 forks source link

Performance Issues on Jetson Xavier NX #230

Open MsWik opened 3 years ago

MsWik commented 3 years ago

CV 4.5.0 + CUDA, jetpack 4.5.1. Low performance of models like yolo4, yolo4x (lower than in the original darcnet), performance of tiny models is correct. Also, when I try to convention any model in INT8, I get the following error: failed to open  /home/andry/test_jetson_xavier/tkDNN-master/tkDNN/src/Int8BatchStream.cpp:69 Aborting...

perseusdg commented 3 years ago

Regarding the int8 model conversion ,have you downloaded the dataset and followed the steps mentioned in the Readme? Cause it looks there is a problem with your dataFilePath for image paths

And about the low performance on Xavier nx ...are you using the current master branch ?

MsWik commented 3 years ago

FP16 / FP32 dimension models are obtained correctly, the error occurs only when I use INT8. yes, the current master branch is used

MsWik commented 3 years ago

Yolo 4, yolo 4x performance issue resolved by reinstalling tkDNN, int8 issue not resolved yet.

mive93 commented 2 years ago

@MsWik have you downloaded the dataset as suggested by @perseusdg ? Have you exported the needed env variables as explained here?

I tested everything yesterday on a Jetson Xavier AGX and I got no problems.