Error when "import torch" is executed after python3

[nagesh_accord]

@dusty_nv

That is generally the case for containers that share the same base image, however you may have containers that don’t share the same base image.

How can we find out which containers share same base image, which wont share?
Any method to find out them.

@dusty_nv

If you are trying to budget what disk space you need to specify for your system, I would err on the side of caution and go bigger than you think you need for future-proofing. 2TB NVME drives are only ~$100 these days, and then you wouldn’t need to worry…

We have 240 GB NVM M2 memory slot currently. Can I select this “NVM…” in the SDK manager instead of “eMMC” to flash my “software components”. Hope it flashs on to NVM memory only and things work fine. Please confirm.

@dusty_nv

Ahh, when I said ‘host device’, I meant your Jetson’s native filesystem outside of docker - not your host PC. The system will be standalone after you flash it from your PC, then no PC needed.

oh ok.
Thanks for the info. So it means we can update the mounted folders from the container terminal itself and these updates will be reflected automatically in the folders in the file system outside the containers.
Basically this can be used for the working folders of the CUDA developers, where they can work inside the container terminal itself and can have access to there code/files etc of the outside file system from here itself. Pls confirm if my understanding is correct.

@dusty_nv

Yes, and those System Setup steps are mostly just a collection of best practices for building large containers onboard your device (for example, changing the docker data-root to NVME so you don’t fill up your smaller eMMC or SD card)

ok fine. Thanks for the information.

@dusty_nv

Yes, they are just apt packages (for example, cuda-toolkit-11-4) and you can uninstall them via apt.

uninstalling via apt commands method, will be from the target. As there are many software components that are installed, uninstalling all of them individually through apt may not workout properly. Some dependency files might remain in some folders itself, that might have side effects with application functionalities later on.

As I will be installing many components of all related to deep learning package, can we uninstall all these fully from the SDK manager from Host PC. Just wanted to know?
Anyhow, I might be opening a new thread seperately for SDK manager installations queries soon.

Thanks.

[dusty_nv]

@nagesh_accord

How can we find out which containers share same base image, which wont share?
Any method to find out them.

You would have to look at their dockerfiles and what base images they used, I don’t think there is a way to realistically determine it. If you are so concerned about the disk space it is probably best to just pull all the containers you need and see what it actually takes.

@nagesh_accord

Can I select this “NVM…” in the SDK manager instead of “eMMC” to flash my “software components”. Hope it flashs on to NVM memory only and things work fine. Please confirm.

I believe so, although I haven’t flashed the OS to NVME before (I tend to flash the OS to eMMC or SD card and use NVME for projects, data, containers, and code). If you encounter problems with flashing to NVME, I recommend you open a new topic about SDK Manager as you alluded to later in your post.

@nagesh_accord

uninstalling via apt commands method, will be from the target. As there are many software components that are installed, uninstalling all of them individually through apt may not workout properly. Some dependency files might remain in some folders itself, that might have side effects with application functionalities later on.

Then I suggest after your initial development, if you end up using containers and don’t need the JetPack packages installed on the device’s native filesystem (outside container), then just re-flash your device again and this time don’t have SDK Manager install them.

[nagesh_accord]

@dusty_nv

I was able to successfully install Nvidia-Jetpack on my device.
Later I was trying to install pytorch, in that when executing the below step, it is giving error as shown below:
Kindly let us know, how to resolve this.

root@trident-desktop:/home/trident# pip3 install Cython<3
bash: 3: No such file or directory

I am following the below steps to install as shared by you in the beginning of the thread:

> Python 3.6

wget https://nvidia.box.com/shared/static/p57jwntv436lfrd78inwl7iml6p13fzh.whl -O torch-1.8.0-cp36-cp36m-linux_aarch64.whl
sudo apt-get install python3-pip libopenblas-base libopenmpi-dev libomp-dev
pip3 install Cython<3
pip3 install numpy torch-1.8.0-cp36-cp36m-linux_aarch64.whl

[dusty_nv]

@nagesh_accord try pip3 install 'Cython<3' instead

[nagesh_accord]

Thanks. It worked for me.

Finally I was able to completely install jetpack 5.1.2 component (with the command “sudo apt-get install nvidia-jetpack”) +pytorch+ opencv built with CUDA on my Jetson Agx Xavier.

Next in case of any further queries or issue, execution of sample programs of CUDa, open cv, torch, onn, tensorrr ,etc from customer will ask here.

Thanks a lot again for the support @dusty_nv

[dusty_nv]

OK great, glad you got it working! For new issues please feel free to open another topic, and consider this one resolved for now.

[nagesh_accord]

Sure. Thanks a lot for your help and guidance. I learnt a lot from you especially this docker, image, container concepts were very new to me.

[nagesh_accord]

@dusty_nv

@nagesh_accord SDK Manager runs on the Ubuntu PC, and I believe it can reset your Jetson into recovery mode without shorting the pins, or if you run sudo reboot –force forced-recovery from your Jetson.

I tried running this command on the target to force the unit in to recovery mode.
it is throwing error saying:

reboot: invalid option – ‘o’

Any idea why?

[dusty_nv]

@nagesh_accord it looks like it needs --force instead of -force

[nagesh_accord]

Yes, It worked.