Installing from conda with official installation command latest PyTorch 1.9.0 installs torchvision 0.2.2 #4240
Comments
vadimkantorov
changed the title
|
@soumith glibc problems are back. I think the problem is that new conda installs python 3.9 and it is not supported There should be a fool-proof way to detect this. This glibc problem isn't infrequent, and happens on all new conda installs. Very strange spec: |
|
This is likely a duplicate of #4076 and/or #4071 @vadimkantorov could you please try #4071 (comment) ? |
|
Again, very strange spec - only Python 2 is supported |
|
Standing questions:
Workaround from #4071 (comment) installs |
Oh, that looks unexpected. Could you run conda in verbose mode and paste the output here? |
PREFIX=/home/lab/lab/vadim/prefix
mkdir -p $PREFIX
curl -L -so $PREFIX/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh && chmod +x $PREFIX/miniconda.sh && $PREFIX/miniconda.sh -b -p $PREFIX/miniconda && rm $PREFIX/miniconda.sh
export PATH=$PREFIX/miniconda/bin:$PATH
conda install -y pytorch torchvision cudatoolkit=11.1 -c pytorch -c nvidia -c conda-forge
conda install -y ffmpeg aria2
#conda install -y av -c conda-forge
pip install opencv-python opencv-contrib-python matplotlib scikit-image imageio wandb visdom sklearn pandas |
|
Also note ffmpeg 4.3 that gets installed which seems not supported. If it's not supported, torchvision package should impose ffmpeg < 4.3 |
|
Thanks for the logs, sorry I brainfarted: torchvision 0.10 is expected, it's our latest stable version https://github.com/pytorch/vision/releases/tag/v0.10.0 |
|
With this installation command ffmpeg does not work: While |
|
I've flagged this to the relevant folks. This is going to be fixed in pytorch 1.9.1 release, and we're working on robust process changes to catch these early. We used to have adhoc pre-release manual testing a long time ago, but we regressed on that aspect, we will go back and implement some of this manual or properly automated testing to catch these. |
|
Tagging @malfet @driazati @zhouzhuojie @seemethere to pull this into our list of issues to investigate. Thanks! |
Conda error reporting could be much improved. There isn't a single torchvision-0.10.0 with such python versions range uploaded to Leaving As conda/conda-build#3956 (comment) suggest, this dep could be stripped during the build time, or could be stripped after package were build by using something along those lines: https://gist.github.com/malfet/932311d0e7321cfdeed9e814a759b84e Uploaded |
|
I've gone through and uploaded the patched binaries for all domains from @malfet's anaconda org and can verify that it resolves the installation issues. Commands used to copyCommands used to verifyNote the channels the Going to close this now that these packages have been updated and should not rely on Feel free to re-open if the issue still persists |
|
The official installation command does not pin the versions. Does it work without version pinning? |
|
While at it, could we also recommend an installation command that installs a workable torchvision + ffmpeg combination? (for both installation and building torchvision from scratch. if torchvision only works with ffmpeg 4.2, it should update its dependency constraints) |
Yes should also work without version pinning as well |
|
@vadimkantorov worth opening a new issue for the ffmpeg thing. Is it that by default, the install commands dont install a working torchvision + ffmpeg? just describe it, and feel free to cc me. |
Yes, as described in this issue, it installs for some reason ffmpeg 4.3 (is 4.3 supported?) from pytorch channel (why?) that misses some libraries |
We will be looking into the issues with ffmpeg soon, thanks for your patience! |
weiji14
added a commit
to weiji14/ctcorenet
that referenced
this issue
Aug 9, 2021
Datasets, Transforms and Models specific to Computer Vision! Using torchvision instead of torchio because the torchio CropOrPad function doesn't work on 2D images, only 3D biomedical ones. Had to downgrade to Python 3.7 because pytorch 1.9.0 and torchvision 0.10.0 had some dependency hell going on recently, see pytorch/vision#4240.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
@soumith @ezyang
My machine is supported, previously with PyTorch 1.8.0, correct torchvision was being installed.
It would be good to somehow test most common conda installation scenarios. These problems have come back so many times. One could run once every major release a matrix with different machines / OS images / GPUs on EC2 just for testing this.
I think it would be good to modify torchvision 0.2.2 binaries so that it shows a message of how to fix it if it's run with recent PyTorch.
Also, for some reason conda website has a torchvision release 0.10.0, while torchvision github has only 0.9.2-rc2.
The text was updated successfully, but these errors were encountered: