Package Details: python-torchvision-cuda 0.6.1-1

Git Clone URL: (read-only, click to copy)
Package Base: python-torchvision
Description: Datasets, transforms, and models specific to computer vision (with GPU support)
Upstream URL:
Licenses: BSD
Conflicts: python-torchvision=0.6.1
Provides: python-torchvision=0.6.1
Submitter: flacks
Maintainer: hottea (lilac)
Last Packager: lilac
Votes: 16
Popularity: 0.45
First Submitted: 2018-01-31 10:46
Last Updated: 2020-06-23 04:46

Dependencies (17)

Required by (18)

Sources (1)

Latest Comments

1 2 3 4 Next › Last »

trougnouf commented on 2020-07-10 20:51

Can you remove 'python-pytorch-cuda' and cuda from the global dependencies?

elanglois commented on 2020-05-02 18:49

Edit: Sorry, I misunderstood split packages and thought it would be possible to separate build(), check() and makedepends within a single PKGBUILD. It's unfortunate that it doesn't seem possible to have variable reuse with useable building for non-CUDA users.

Original comment:

@hottea You don't have to provide two PKGBUILDs but the code for build() and check() is already split into separate commands for the gpu and cpu versions so why not separate those into different function names so that no large unnecessary packages have to be installed for cpu users? Regardless of whether its possible to download the package elsewhere if this PKGBUILD is claiming to provide python-torchvision (which would be mainly used by non-CUDA users) then I think it makes sense to allow building python-torchvision without CUDA.

hottea commented on 2020-05-02 09:00

@elanglois You don't need to build it yourself. You could install it from ArchLinux CN repo. I have no plan to provide two PKGBUILD for pytorch-torchvision and python-torchvision-cuda.

elanglois commented on 2020-05-02 02:34

Could you split up build(), check(), and makedepends into separate functions for cuda and non-cuda versions? python-torch-cuda conflicts with python-torch so using this PKGBUILD to build python-torchvision means uninstalling python-torch, installing python-torch-cuda, building python-torchvision, then uninstalling python-torch-cuda and reinstalling python-torch. Sure I could modify the PKGBUILD (and have been) but it doesn't make sense to be providing python-torchvision if the existing PKGBUILD doesn't work well for it.

~~Really, I think it would be best just to split this into separate PKGBUILDS for python-torchvision and python-torchvision-cuda but splitting the remaining functions works too.~~ [edit: not sure about whether I think separate PKGBUILDS would be preferable, rest still stands]

hottea commented on 2020-04-25 12:43

@promach Maybe you're running out of memory. python-torchvision and python- torchvision-cuda are built with a single PKGBUILD, but with different options. Check the wiki if you have question about PKGBUILD.

promach commented on 2020-04-25 09:22

May I know why the git clone url is the same for non-cuda version of python-torchvision ?

And I face some compilation errors during building the library.

test/ PASSED                                                          [ 31%]
test/ PASSED                                              [ 32%]
test/ PASSED                                               [ 32%]
test/ PASSED                                                      [ 32%]
test/ PASSED                                                      [ 33%]
test/ PASSED                                                      [ 33%]
test/ PASSED                                                      [ 34%]
test/ /home/phung/Downloads/python-torchvision/PKGBUILD: line 60: 17596 Killed                  PYTHONPATH="${PWD}/build/lib.linux-${CARCH}-$(get_pyver)" pytest -v
==> ERROR: A failure occurred in check().

hottea commented on 2020-04-22 15:23

@jclds139 You could modify this PKGBUILD as you need. Modifications you need include: 1. remove python-torchvision-cuda from pkgname. 2. remove python-pytorch-cuda from makedepends. 3. remove all lines related to build or package python-torchvision-cuda

Another solution, just install the binary package from ArchLinxCN repo.

jclds139 commented on 2020-04-21 16:49

I can see why you have python-pytorch-cuda included as a dependency, but I still have to disagree with your reasoning. Since this package isn't going to use CUDA, it shouldn't be built with it or the FORCE_CUDA=1 option. All that's doing is forcing us to download the huge cuda package that we don't need in order to build this. Is there any way you could split this PKGBUILD so that this package doesn't require CUDA in order to build?

hottea commented on 2020-04-16 13:31

@horst python-torchvision requires python-pytorch, python-torchvision-cuda requires python-pytorch-cuda. I add python-pytorch-cuda to makedepends, because I need to build torchvision in a clean chroot environment.

horst commented on 2020-04-16 13:14

With python-torchvision-cuda existing, should python-torchvision require python-pytorch-cuda in addition to python-pytorch?