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

build tensorRT engine on jetson nano get ERROR! #193

Closed jcyhcs closed 2 years ago

jcyhcs commented 3 years ago

i download yolo4-csp weights, then i use test_yolo4-csp for convert to TensorRT engine,but got error: look this picture: image please help!

MahmoudF98 commented 3 years ago

+1

rickymedrano commented 3 years ago

Tested it and got same error on the Nano. Not sure if this could be the problem, but yolo4-csp uses logistic activation in a couple layers and it doesn't show logistic as one of the supported activations https://github.com/ceccocats/tkDNN#darknet-parser

mive93 commented 3 years ago

I have just tested that and it works for me. Given that I have changed those with my last commit (because the original model changed) be sure to have pulled the latest version, deleted the old weight folder and the old .rt file and create a new one just running ./test_yolo4-csp which automatically downloads the corrects weights and create the correct rt.

rickymedrano commented 3 years ago

@mive93 I was using the latest version of this repo and was still getting the error. @MahmoudF98 @jcyhcs Figured it out. It was being killed because of insufficient memory. The 2GB swap, used by default on Jetpack 4.5, wasn't enough. You need to increase it to 4GB. Once I set up 4GB swap, with instructions from here, I was able to create the .rt file successfully for yolo4-csp.

mive93 commented 3 years ago

Oh that's right, I forgot it was on the Nano. Then in case you can follow the instructions given here: https://github.com/ceccocats/tkDNN/issues/26. Indeed, the problem with the Nano is the memory.

mive93 commented 2 years ago

Closing for inactivity. Feel free to reopen.