No module named 'torch._c'

[marcusvlc]

Im trying to install torch on raspberry pi 3 B+ with your .whl files, but im gotting the following error:

No module named 'torch._c'

pytorch installs normally, the error comes up when I try to import

I searched on the internet but found nothing. Can you help me?

Im using raspibian lite, python3.7 and i alredy installed all libraries.

I tried to install the torch-1.1.0 cp37 armv7l

[nmilosev]

Raspberry Pi 3 should be ARM64, so you should use aarch64 build?

[marcusvlc]

I was unable to install the armv6l and aarch64 files because it says they are not compatible with this platform. The only one I could install was armv7l but I got the error I described above.

Obs -- Raspbiain lite (my current system) is armv7l

[nmilosev]

Could you please provide the output of uname -a? Thanks!

[marcusvlc]

uname -a on raspibian lite returns armv7l.

[nmilosev]

Please see the discussion in #3 about this issue. I think you need an actual 64-bit OS to run pytorch for 64 bits.

[marcusvlc]

I also tried to install the aarch version on aarch64 ubuntu mate 64 bits for raspberry pi. Im getting error with a version of a library, when i try to import the torch

ImportError: /usr/lib/x86_64-linux-gnu/libstdc++.so.6: version `GLIBCXX_3.4.26' not found

[radiodee1]

ImportError: /usr/lib/x86_64-linux-gnu/libstdc++.so.6: version `GLIBCXX_3.4.26' not found

I get this error too. I'm using raspberrypi Buster. It seems libstdc++.so.6 in FC30 might be ahead of Buster on the raspberry pi. What could I do about this??

[nmilosev]

Okay, thank you for the info. Really means a lot.

I checked my glibc version and indeed it is newer:

Name         : glibc
Version      : 2.29
Release      : 15.fc30
Architecture : x86_64
Size         : 25 M
Source       : glibc-2.29-15.fc30.src.rpm
Repository   : @System
From repo    : updates

What could I do about this??

Sadly not much. I don't think updating glibc is easy, as it would break the rest of your system.

I just had a thought, what about running a chroot? I have a tool made for that called anyfed (https://github.com/nmilosev/anyfed) or you could use Docker? That could work but I can't guarantee. Anyfed hasn't been updated in a while so I will update it today if you want to try it.

Of course it would be best to get Fedora on your device and use the builds that way. :)

https://wiki.debian.org/Debootstrap has an option for different architecture, which means that in theory you could make a chroot with Debian in it and compile Pytorch there, similar to how I am doing it with Fedora.

[radiodee1]

I'm trying to follow your instructions but use fc29. It might be a dumb idea but the build isn't finished yet. I'll tell you how it goes.

…

On Mon, Sep 2, 2019, 11:06 AM Nemanja Milosevic ***@***.***> wrote: Okay, thank you for the info. Really means a lot. I checked my glibc version and indeed it is newer: Name : glibc Version : 2.29 Release : 15.fc30 Architecture : x86_64 Size : 25 M Source : glibc-2.29-15.fc30.src.rpm Repository : @System From repo : updates What could I do about this?? Sadly not much. I don't think updating glibc is easy, as it would break the rest of your system. I just had a thought, what about running a chroot? I have a tool made for that called anyfed (https://github.com/nmilosev/anyfed) or you could use Docker? That could work but I can't guarantee. Anyfed hasn't been updated in a while so I will update it today if you want to try it. Of course it would be best to get Fedora on your device and use the builds that way. :) https://wiki.debian.org/Debootstrap has an option for different architecture, which means that in theory you could make a chroot with Debian in it and compile Pytorch there, similar to how I am doing it with Fedora. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACB527AJYBTAVKJITBWMFHTQHUTYTA5CNFSM4IOV7RX2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5WB2PY#issuecomment-527179071>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACB527EC6TJUJVRKW3RJIQLQHUTYTANCNFSM4IOV7RXQ> .

[nmilosev]

Okay, great! :)

Anyfed just updated to Fedora 30.

Cheers

[marcusvlc]

@radiodee1 If all goes well, could you share the compiled pytorch and torchvision files? And also tell which system managed to install them.

[radiodee1]

I am interested in running a torch model on the raspberry pi 4. It has python 3.7

I got a virtualbox instance and I installed FC29 on it. I run ubuntu.

I followed the instructions for building the torch wheel, but I used FC29 everywhere that FC30 was specified. That's first. That means the one-liner in the instructions should read as follows:

sudo dnf install --releasever=29 --installroot=/tmp/F29ARM --forcearch=armv7hl --repo=fedora --repo=updates systemd passwd dnf fedora-release vim-minimal openblas-devel blas-devel m4 cmake python3-Cython python3-devel python3-yaml python3-setuptools python3-numpy python3-cffi python3-wheel gcc-c++ tar gcc git make tmux -y

After that I just followed the instructions. This took about 10 hours. Then I move the torch wheel to the pi and install with pip3.

Then, you need to follow the instructions here: pytorch/pytorch#574 . Basically you need to change the names of _C and _dl after you install the library with pip3.

After that it seems to work. I never tried torchvision.

I hesitate to share my wheel because technically it didn't work. You still have to modify the _C and _dl files after you install.

[nmilosev]

This is great! So this is the possible solution to this problem, albeit manual. Still, better than building from source.

Thanks for your report. @marcusvlc Could you follow the instructions in the linked issue and try again?

[nmilosev]

@radiodee1 Please share your whl so we can try it out.

[radiodee1]

how do i share my wheel?

[marcusvlc]

@radiodee1 Thanks for you aswner, i will try it. Can you put you wheel file on a git repository (or google drive) and share with us?

[nmilosev]

how do i share my wheel?

It would be best to attach it here in the issues?

[nmilosev]

@radiodee1 Thanks for you aswner, i will try it. Can you put you wheel file on a git repository (or google drive) and share with us?

Please also try the original file for armv7?

[radiodee1]

I forked the repo and changed/added my whl file. See here: https://github.com/radiodee1/pytorch-arm-builds . Funny thing is, the author of the commit was recorded as someone else, not me.

[runninbear5]

@radiodee1

I have been trying to make my own .whl files similar to what you have done and I had a lot of trouble. I see here you said

I got a virtualbox instance and I installed FC29 on it. I run ubuntu.

I followed the instructions for building the torch wheel, but I used FC29 everywhere that FC30 was specified. That's first. That means the one-liner in the instructions should read as follows:

When you say you run ubuntu does that mean the machine you were running the vm on was an ubuntu machine? When I was struggling to get fedora running I was using a windows machine.

Could you document how you when about setting up the virtual machine to help in the future?

[nmilosev]

Running Fedora in a VM for this purpose would not work. You need to virtuelize ARM with qemu and nested virtualization is tricky to say at least. You have to run on bare hardware.

[radiodee1]

Though I would agree that nested virtualization is not generally a good idea, I want to note that that is exactly what I did and I ended up with a wheel I could use. I don't recommend this for others but I wanted to note that that was my method. So I used ubuntu as my native operating system. I used virtualbox for my fist virtualization, and I used FC29 for the guest inside virtualbox. Then I used FC29 as the nested os for arm. I wanted to be really clear.