Jetson flashed with Jetpack 6.0 DP using sdkmanager. Username: jetson Password: uas-2024
Jetson requires a different torch wheel than the PyPi one to use GPU, which is achieved with the dustynv/l4t-pytorch:r36.2.0 docker image. A modified installation of ultralytics is also required to use this installation of torch. Here we reference the official ultralytics Dockerfile for jetson devices. To sync this up with the normal branch of development (for everyone not developing on the jetson), the original Dockerfile has been updated to use the ultralytics/ultralytics:8.1.10 image.
The official Loki plugin for Docker doesn't work for ARM, so an alternative has been provided, which can be installed with make build-arm
Version info:
Jetpack 6.0 - CUDA 12.2.12, cuDNN 8.9.4, TensorRT 8.6.2, Ubuntu-based file system
Overview:
Jetson flashed with Jetpack 6.0 DP using sdkmanager. Username: jetson Password: uas-2024
Jetson requires a different
torchwheel than the PyPi one to use GPU, which is achieved with thedustynv/l4t-pytorch:r36.2.0docker image. A modified installation ofultralyticsis also required to use this installation oftorch. Here we reference the official ultralytics Dockerfile for jetson devices. To sync this up with the normal branch of development (for everyone not developing on the jetson), the original Dockerfile has been updated to use theultralytics/ultralytics:8.1.10image.The official Loki plugin for Docker doesn't work for ARM, so an alternative has been provided, which can be installed with
make build-armVersion info:
Usage Notes:
To use GPU in Docker Containers:
docker-compose.ymlIf running on the Jetson, change the dockerfile on line 5 of
docker-compose.ymltoDockerfile-jetsonTo ssh into the jetson, connect via usb-c or ethernet and run
ssh jetson@192.168.55.1For performance (seems to help a little bit):
sudo nvpmodel -m 0sudo jetson_clocks